Professional Documents
Culture Documents
Leonard Barolli
Makoto Takizawa
Fatos Xhafa
Tomoya Enokido Editors
Web, Artificial
Intelligence
and Network
Applications
Proceedings of the Workshops of the
33rd International Conference on
Advanced Information Networking
and Applications (WAINA-2019)
Advances in Intelligent Systems and Computing
Volume 927
Series editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
The series “Advances in Intelligent Systems and Computing” contains publications on theory,
applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all
disciplines such as engineering, natural sciences, computer and information science, ICT, economics,
business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the
areas of modern intelligent systems and computing such as: computational intelligence, soft computing
including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms,
social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and
society, cognitive science and systems, Perception and Vision, DNA and immune based systems,
self-organizing and adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics including
human-machine teaming, knowledge-based paradigms, learning paradigms, machine ethics, intelligent
data analysis, knowledge management, intelligent agents, intelligent decision making and support,
intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings
of important conferences, symposia and congresses. They cover significant recent developments in the
field, both of a foundational and applicable character. An important characteristic feature of the series is
the short publication time and world-wide distribution. This permits a rapid and broad dissemination of
research results.
** Indexing: The books of this series are submitted to ISI Proceedings, EI-Compendex, DBLP,
SCOPUS, Google Scholar and Springerlink **
Advisory Board
Chairman
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
e-mail: nikhil@isical.ac.in
Members
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa
Clara, Cuba
e-mail: rbellop@uclv.edu.cu
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: escorchado@usal.es
Hani Hagras, School of Computer Science & Electronic Engineering, University of Essex, Colchester, UK
e-mail: hani@essex.ac.uk
László T. Kóczy, Department of Information Technology, Faculty of Engineering Sciences, Győr, Hungary
e-mail: koczy@sze.hu
Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA
e-mail: vladik@utep.edu
Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan
e-mail: ctlin@mail.nctu.edu.tw
Jie Lu, Faculty of Engineering and Information, University of Technology Sydney, Sydney, NSW, Australia
e-mail: Jie.Lu@uts.edu.au
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: epmelin@hafsamx.org
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: nadia@eng.uerj.br
Ngoc Thanh Nguyen, Wrocław University of Technology, Wrocław, Poland
e-mail: Ngoc-Thanh.Nguyen@pwr.edu.pl
Jun Wang, Department of Mechanical and Automation, The Chinese University of Hong Kong, Shatin,
Hong Kong
e-mail: jwang@mae.cuhk.edu.hk
Editors
123
Editors
Leonard Barolli Makoto Takizawa
Department of Information Department of Advanced Sciences
and Communication Engineering Hosei University
Fukuoka Institute of Technology Koganei-Shi, Tokyo, Japan
Fukuoka, Japan
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Welcome Message from AINA-2019 Steering
Committee Co-chairs
v
vi Welcome Message from AINA-2019 Steering Committee Co-chairs
Makoto Takizawa
Leonard Barolli
AINA Steering Committee Co-chairs
Welcome Message from AINA-2019 General
Co-chairs
It is our great pleasure to welcome you all at the 33rd International Conference on
Advanced Information Networking and Applications (AINA-2019), which will be
held at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
AINA International Conference is a forum for sharing ideas and research work in
the emerging areas of information networking and their applications. The area of
advanced networking has grown very rapidly, and the applications around it have
experienced an explosive growth especially in the area of pervasive and mobile
applications, sensor networks, ad hoc networks, vehicular networks, multimedia
computing and social networking, semantic collaborative systems, as well as grid,
P2P, IoT, and cloud computing. This advanced networking revolution is trans-
forming the way people live, work, and interact with each other, and is impacting
the way business, education, entertainment, and health care are operating. The
papers included in the proceedings cover theory, design, and application of
computer networks, distributed computing, and information systems.
Each year AINA receives a lot of paper submissions from all around the world. It
has maintained high-quality accepted papers and is aspiring to be one of the main
international conferences on the information networking in the world. In
conjunction with AINA-2019 conference, there are 14 workshops, which also
accepted good-quality papers.
An international conference of this size requires the support and help of many
people. A lot of people have helped and worked hard to produce a successful
AINA-2019 technical program and conference proceedings. First, we would like to
thank all authors for submitting their papers, the session chairs, and distinguished
keynote speakers. We are indebted to program area chairs, Program Committee
Members and Reviewers, who carried out the most difficult work of carefully
evaluating the submitted papers.
We would like to give our special thanks to Prof. Makoto Takizawa and Prof.
Leonard Barolli, the Co-chairs of the Steering Committee for their strong encour-
agement, guidance, and insights, and for spending a lot of energy for conference
organization and shaping the conference program. We would like to thank PC
Co-chairs and Workshop Co-chairs of AINA-2019 for their great contribution to the
vii
viii Welcome Message from AINA-2019 General Co-chairs
success of the conference. Our special thanks go to the Finance Chair and Web
Administrator Co-chairs.
Finally, we would like to thank the Local Arrangement Team for the support and
good local arrangement for the conference. We do hope that you will have a great
time in Matsue, Japan.
Tomoya Enokido
Farookh Hussain
Alireza Shahrabi
AINA-2019 General Co-chairs
Welcome Message from AINA-2019 Program
Committee Co-chairs
ix
x Welcome Message from AINA-2019 Program Committee Co-chairs
workshops and symposiums, which enrich the conference and provide additional
opportunities for discussions and future cooperations.
The great success of the AINA conference series would not be possible without
the enormous commitment and support of the Steering Committee co-chairs Prof.
Makoto Takizawa and Prof. Leonard Barolli. Therefore, we would like to thank
them for their strong encouragement and guidance.
The general coordination of an event such as AINA conference requires a lot of
coordination effort as well as many other activities related to the conference
organization. Here, we thank the General Co-chairs for their great support and
invaluable suggestions. We give special thanks to the Finance Chair and Web
Administrator Co-chairs for their great efforts and efficient work to deal with many
conference matters.
We hope you will enjoy the conference and readings and have a great time in
Matsue, Japan.
Akimitsu Kanzaki
Flora Amato
Omar Hussain
AINA-2019 Program Committee Co-chairs
Welcome Message from AINA-2019 Workshops’
Co-chairs
xi
xii Welcome Message from AINA-2019 Workshops’ Co-chairs
11. The 4th International Workshop on Big Data Processing in Online Social
Network (BOSON-2019)
12. The 2nd International Workshop on Internet of Everything and Machine
Learning Applications (IOEMLA-2019)
13. The 1st International Workshop on Multi-Clouds and Mobile Edge Computing
(M2EC-2019)
14. The 1st Workshop on Artificial Intelligence and Machine Learning
(AIMAL-2019)
We would like to thank the community for their great response to AINA-2019
workshops. The excellent technical program of the workshops was the result of a
professional work from Workshop Chairs, workshop Program Committees,
Reviewers, and authors.
We would like to give our special thanks to Prof. Makoto Takizawa and
Prof. Leonard Barolli, the Steering Committee Chairs of AINA International
Conference, for their strong encouragement and guidance to organize the
AINA-2019 workshops and symposiums. We would like to thank AINA-2019
General Co-chairs their advices to make possible organization of AINA-2019
workshops and symposiums. We are thankful to AINA-2019 Program Co-chairs for
their support and help to prepare the technical program of AINA-2019 workshops
and symposiums.
We wish all of you entertaining and rewarding experience in all workshops and
AINA-2019 International Conference.
Hui-Huang Hsu
Omid Ameri Sianaki
Rubem Pereira
AINA-2019 Workshop Co-chairs
Welcome Message from AINA-2019 Workshops’ Co-chairs xiii
General Co-chairs
Workshop Co-chairs
Award Co-chairs
Publicity Co-chairs
Finance Chair
Web Chairs
xv
xvi Welcome Message from FINA-2019 Symposium Co-chairs
Akimitsu Kanzaki
Flora Amato
Omar Hussain
FINA-2019 Symposium Co-chairs
Welcome Message from FINA-2019 Symposium Co-chairs xvii
Leonard Barolli
Arjan Durresi
HWISE-2019 International Workshop Co-chairs
xix
xx Welcome Message from HWISE-2019 International Workshop Co-chairs
Workshop Co-chairs
Advisory Co-chairs
Kin Fun Li
Kosuke Takano
MAW-2019 Symposium Co-chairs
xxi
xxii Welcome Message from MAW-2019 International Symposium Co-chairs
Leonard Barolli
Arjan Durresi
BICom-2019 Co-chairs
xxiii
xxiv Welcome Message from BICom-2019 International Workshop Co-chairs
Workshop Co-chairs
Advisory Co-chairs
Rubem Pereira
Jeu-Yih Jeng
TeNAS-2019 General Chairs
Teh-Sheng Huang
Shin-Hao Chang
TeNAS-2019 Program Committee Co-chairs
xxv
xxvi Welcome Message from TeNAS-2019 Workshop Organizers
General Co-chairs
Program Co-chairs
Yoshitaka Shibata
Noriki Uchida
IWDENS-2019 Workshop Co-chairs
xxix
xxx Welcome Message from IWDENS-2019 Workshop Organizers
Leonard Barolli
COLLABES-2019 Organizer
xxxi
xxxii Welcome Message from COLLABES-2019 International Workshop Organizer
Anne Kayem
INTRICATE-SEC-2019 Workshop Chair
xxxiii
xxxiv Welcome Message from INTRICATE-SEC-2019 Workshop Chair
Welcome to the 5th on Engineering Energy Efficient Internet Worked Smart seNsor
(E3WSN-2019), which is held in conjunction with the 33rd International
Conference on Advanced Information Networking and Applications (AINA-2019)
at Kunibiki Messe, Matsue, Japan, from March 27–29, 2019.
This workshop aims at bringing together academics and practitioners from dif-
ferent areas (such as software engineering, system engineering, energy efficiency,
and embedded systems) to promote design, validation, and implementation of
energy-efficient inter-networked smart sensors. The papers included in the pro-
ceedings cover novel languages, designs, simulation, and applications of
inter-networked smart sensors.
E3WSN-2019 contains high-quality research papers submitted by researchers
from all over the world. Each submitted paper was peer-reviewed by Reviewers
who are experts in the subject area of the paper. Based on the review results, the
Program Committee accepted good-quality papers.
For organizing an international workshop, the support and help of many people
are needed. First, we would like to thank all authors for submitting their papers. We
also appreciate the support from Program Committee members and Reviewers who
carried out the most difficult work of carefully evaluating the submitted papers.
The E3WSN organization team would like to give its special thanks to the
Steering Committee Chairs, General Co-chairs, Program Co-chairs, and Workshop
Organizers for their support to organize the workshop.
Leonardo Mostarda
Jonathan Loo
E3WSN-2019 Co-chairs
xxxv
xxxvi Welcome Message from E3WSN-2019 International Workshop Co-chairs
Information and networking technologies have been a major driving force for the
advances of human society for decades. This 4th edition of International Workshop
on Innovative Technologies in Informatics and Networking (WITIN-2019) aims to
attract innovative attempts in the related fields. This workshop will provide a
platform for researchers and practitioners to meet and exchange their ideas.
WITIN-2019 will be held in conjunction with the 33rd International Conference
on Advanced Information Networking and Applications (AINA-2019) at Kunibiki
Messe, Matsue, Japan, from March 27–29, 2019. We accepted very good-quality
papers for WITIN-2019 after a rigorous selection process. We hope all of you will
enjoy the workshop and find it a great opportunity to discuss research works.
Many people contributed to the CFP and paper review of WITIN-2019. We wish
to thank the Program Committee members for their excellent effort in screening and
selecting the papers. We would like to express our gratitude to Prof. Makoto
Takizawa and Prof. Leonard Barolli, the Steering Committee Co-chairs of AINA
International Conference, for their encouragement and guidance. We are very
thankful to General Co-chairs, PC Co-chairs, and Workshop Co-chairs of
AINA-2019, for their great organization of this conference. Last but not least, we
would like to thank and congratulate all the authors for their support to the
workshop.
Hui-Huang Hsu
Yi-Jen Su
WITIN-2019 Workshop Co-chairs
xxxvii
xxxviii Welcome Message from WITIN-2019 International Workshop Organizers
Workshop Co-chairs
xxxix
xl Welcome Message from BOSON-2019 International Workshop Chair
and guidance to organize the workshop. I would like to thank AINA-2019 General
Co-chairs, PC Co-chairs, and Workshop Co-chairs for their support and help to
prepare the technical program.
I hope you will enjoy BOSON workshop and AINA International Conference,
find this a productive opportunity for sharing experiences, ideas, and research
results with many researchers, and have a great time in Matsue, Japan.
Flora Amato
BOSON-2019 Workshop Chair
Welcome Message from BOSON-2019 International Workshop Chair xli
xliii
xliv Welcome Message from IOEMLA-2019 International Workshop Co-chair
xlv
xlvi Welcome Message from M2EC-2019 Workshop Organizers
Dimitris Apostolou
Thomas Dreibholz
Feroz Zahid
M2EC-2019 Workshop General Co-chairs
Welcome Message from M2EC-2019 Workshop Organizers xlvii
Publicity Chairs
Nadeem Javaid
AIMAL-2019 Workshop Chair
xlix
l Welcome Message from AIMAL-2019 Workshop Chair
Markus Aleksy
liii
Lévy Walk on Graphs: Message Dissemination
and Uninformed Search
Naohiro Hayashibara
lv
Contents
lvii
lviii Contents
1 Introduction
Recently the increased capability of modern radar systems has enabled the use
of very complex and effective signal processing. This has allowed the radar to
successfully adapt the transmit waveform and receive filter to the operating envi-
ronment in a very sophisticated manner. One attractive problem surely consists
in waveform design for extended targets motivated by the high potentials of
modern High-Resolution Radar (HRR) systems. Moreover, polarization diver-
sity can potentially improve the radar performance w.r.t. conventional radar
systems. Therefore, based on the aforementioned considerations, in this paper
we consider the problem of jointly optimizing the transmit waveform and receive
filter of polarimetric radars for the extended target in clutter.
2 System Model
Assuming a monostatic HRR being employed, the overall scattering behavior of
the probed target is completely specified by the TIRM. It is well known that the
Target Impulse Response Matrix (TIRM) is strongly dependent on the TAA [5],
namely the target scattering behavior is sensitive with respect to even a small
change of the TAA θ. Therefore, taking into account TIRM variability w.r.t.
1
We adopt the notation RN , CN and HN respectively for the sets of N -dimensional
vectors of real numbers, the sets of N -dimensional vectors of complex numbers and
N ×N Hermitian matrices. CN ×N denotes N ×N complex matrices. We use boldface
for vectors a (lower case) and matrices A (upper case). The curled inequality symbol
(and its strict form ) is used to denote generalized matrix inequality: for any
A ∈ HN , A 0 means that A is a positive semidefinite matrix (A 0 for positive
definiteness). The notation A(i, j) represents the i-th row and j-th column element
of the matrix A. I and 0 denote, respectively, the identity matrix and the matrix
(or vector) of zero entries, with their dimensions being determined from the context.
The Euclidean norm of the vector x is denoted by ||x || and for the matrix B,
B2 and BF represent its induced 2-norm and Frobenius norm, respectively. The
symbols (·)∗ , (·)T , (·)† , E[·], (·) and ⊗ denote the conjugate, transpose, Hermitian,
expectation, real part and Kronecker product operators, respectively. Finally, for any
optimization problem P, v(P) and Ω(P) denote the corresponding optimal value and
feasible set, respectively.
Adaptive Waveform Design of Polarimetric Radar 5
where the subscripts H and V denote the horizontal and vertical polarization
components, respectively.
Let the support interval of the TIRM (corresponding to the TAA θ) have
length Q and the observations be collected over a discrete interval M , where
M = Q + N − 1 so that the majority of the extended target echoes is received.
For a certain TAA θ and range bin n, the full-polarization target scattering
matrix T n (θ) ∈ C2×2 , being a matrix-valued sample of the TIRM T (θ), is
expressed as
T (θ) THV,n (θ)
T n (θ) HH,n , (2)
TV H,n (θ) TV V,n (θ)
where θ ∈ [0, 2π), n ∈ {0, 1, . . . , Q − 1}, and TXY,n denotes the scattering coeffi-
cient experienced by transmitted “X” polarization component received on “Y ”
polarization at n-th range bin. Similarly, the clutter scattering matrix associated
to the n-th range bin, being an ensemble of the Clutter Impulse Response Matrix
(CIRM) C , is defined as
CHH,n CHV,n
Cn , n ∈ {−N + 1, . . . , M − 1}. (3)
CV H,n CV V,n
where αT is a complex parameter that takes into account radar equation, prop-
agation effects, etc, and v is the 2M -dimensional additive noise
T
v vH,0 , vV,0 , . . . , vH,M −1 , vV,M −1 ∈ C2M . (5)
2
Generally, it is hard to obtain the exact knowledge of target scattering behavior
before the real illumination. However, by employing a cognitive radar theory, it is
reasonable to suppose that the radar system has access to a (potentially dynamic)
database containing a priori electromagnetic reflectivity of the target. Meanwhile,
the TAA can also be considered as a prior known by estimating the target attitude
to the radar.
6 M. Yuan et al.
Q−1 M −1
The matrices T (θ) n=0 J n ⊗ T n (θ) and C n=−N +1 J n ⊗ C n are the
2M × 2N -dimensional TIRM and CIRM, respectively, with J n being the M ×
1 if 1 − 2 = n
N -dimensional shift matrix defined as J n (1 , 2 ) , 1 ∈
0 if 1 − 2 = n
{1, . . . , M }, 2 ∈ {1, . . . , N }. In the following, we will assume that the clutter
and noise are uncorrelated. Let the noise vector v be zero-mean (i.e., E[v ] =
0), and circularly symmetric with covariance matrix Σv = E[v v † ] = σv2 I . Let
CHV,n = CV H,n due to the reciprocity of monostatic scenario. We do not make
any (restrictive) assumption on the multivariate statistical characterization of
c. Differently, for the clutter scatterer in the n-th range bin, assume it to be
zero-mean, i.e., E[C n ] = 0 (viz. E[C ] = 0). Also, we account for the clutter
correlation in both polarization and range domains, namely define
∗
E[Cj,n Ck,n ]
rjk (n, n )
, (6)
E[|Cj,n | ] E[|Ck,n |2 ]
2
3 Problem Formulation
After filtering the observation vector r through w [w(0), . . . , w(2M − 1)]T ∈
C2M , the resulting system output can be written as
y = w † r = αT w † T (θ)s + w † C s + w † v , (7)
|αT |2 |w † T (θ)s|2
SINRθ (s, w ) , (8)
E[|w † C s|2 ] + E[|w † v |2 ]
which can be equivalently recast as:
|αT |2 |w † T (θ)s|2
SINRθ (s, w ) = . (9)
w † (Σc (s) + σv2 I ) w
We consider the output SINR as the relevant figure of merit. Moreover, to
control some relevant features of the probing signal (such as range-Doppler res-
olution, variations in the signal modulus, and peak sidelobe level and etc.), in
addition to a constraint on the transmitted energy, i.e., ||s||2 = 1, a similar-
ity constraint (||s − s 0 ||2 ≤ γ) is enforced on the transmit sequence where the
3
The explicit form of the clutter covariance matrix is not shown due to reasons of
space.
Adaptive Waveform Design of Polarimetric Radar 7
parameter 0 ≤ γ < 1 rules the size of the trust hypervolume and s 0 is a prefixed
code (||s 0 ||2 = 1). Thus, the optimization problem for the joint design of trans-
mit code and receive filter of full-polarization radar, under the aforementioned
energy and similarity constraints, can be formulated as
⎧
⎪ |w † T (θ)s|2
⎨ max
s,w w † Σ (s)w +σ 2 w † w
c v
P s.t. s 2
=1 , (10)
⎪
⎩
s − s 0 ≤ γ
2
in which the term |αT |2 is not included since it does not affect the optimization
problem being investigated.
|w † T (θ)s|2
g(s, w) max . (11)
s,w w † Σc (s)w + σv2 w † w
and
|w † T (θ)s (m) |2
Pw (m) max w † Σ c (s (m) )w +σv2 w † w
. (13)
w
Proof. The proof is similar to that found in [2] and thus it is omitted for the
sake of brevity.
(m)
Then, P̄w is the well-known minimum-variance distorsionless response
(MVDR) [8] problem and its solution is the Capon filter explicitly as
−1
Σc s (m) + σv2 s (m)2 I T (θ) s (m)
w (m) = −1/2 2 . (15)
Σc s (m)
+ σv s I
2 (m) 2
T (θ) s (m)
|w† T(θ)s|2
max (16)
s,w s† Σ̄c (w) + σv2 w† wI s
where Σ̄c (w) represents the filter-dependent clutter covariance matrix and its
explicit expression is omitted for the sake of brevity.
Proof. The proof is similar to that found in [10] and thus it is omitted for the
sake of brevity.
This is a fractional quadratic problem and to solve it, we follow the guidelines
in [2]. Introducing the parameters
†
W = w (m−1)† T (θ) w (m−1)† T (θ) , (18)
and
M = Σ̄c∗ (w (m−1) ) + σv2 w (m−1)† w (m−1) I , (19)
Adaptive Waveform Design of Polarimetric Radar 9
Problem P2 and P2 are clearly equivalent.
Moreover, by dropping the rank-one constraint of P2 , we obtain the semidef-
inite programming (SDP) as problem P3 :
⎧
⎪
⎪ tr (Q 1 S )
⎪
⎪ max
⎪
⎪ S tr (Q 0 S )
⎨
s.t. tr (Q 1 S ) ≤ 0
P3 , (21)
⎪
⎪ tr (Q 2 S ) = 1
⎪
⎪
⎪
⎪ tr (Q 3 S ) = 1
⎩
S 0
and
I 0 00
Q2 = , Q3 = . (23)
00 01
As a consequence, the optimal solution s (m) of P2 (or equivalently of P2 ) can
be found by employing a semi-definite relaxation (SDR) procedure [9].
5 Numerical Results
In this section, the performance analysis of the proposed algorithm is presented.
We utilize the Georgia Technology Research Institute publicly released dataset
[1] as the foreknowledge of the target. The dataset includes a turntable phase
history data of the T-72 tank measured covering a set of azimuth and elevation
angles by a fully polarimetric X-band radar. More detailed description of the
experimental setup is omitted for the sake of simplicity in this summary and
will be brought in the full version.
(a) (b)
(c) (d)
Fig. 1. The optimized SINR behavior vs. iterations for γ = [0.01, 0.10, 0.20, 0.50]:
(1) (2) (3) (4)
(a) θaz = 27.00◦ ; (b) θaz = 48.70◦ ; (c) θaz = 91.20◦ ; (d) θaz = 133.70◦ .
|αT |2 |w † T (θ)s |2
SINR(θ) . (24)
E[|w † C s |2 ] + E[|w † v |2 ]
Obviously, the case θ = θ0 means the real TAA mismatches with the nominal
one, and then a SINR loss may appear in comparison of its value at the position
of θ = θ0 .
Let the TAA uncertainty interval be θ ∈ [25.5◦ , 28.7◦ ] and the nominal TAA
θ0 = 27.0◦ . Figure 2 depicts SINR(θ) for γ = 0, 0.05, 0.1 and 0.5, respectively.
The curves highlight that, the proposed optimization approach (γ = 0.05, 0.1
and 0.5) achieves a significant SINR improvement w.r.t the fixed transmit wave-
form design (γ = 0). Furthermore, even though the optimization design shows
the maximum SINR is obtained at the position θ = θ0 , it exhibits higher and
12 M. Yuan et al.
Fig. 2. SINR(θ) vs. θ in the TAA uncertainty interval θ ∈ [25.5◦ , 28.7◦ ], with the
nominal TAA θ0 = 27.0◦ and γ ∈ {0, 0.05, 0.1, 0.5}.
flatter SINR values over the entire TAA uncertainty set w.r.t the fixed transmit
waveform design, which shows its robustness.
6 Conclusions
Joint optimization of the transmit signal and receive filter for the extended
target and polarimetric radar in the presence of signal-dependent and signal-
independent interference, has been considered. By considering the SINR at the
output of a filter as the relevant figure of merit for optimization, a similarity
constraint has been enforced on the transmit code, in addition to the energy
constraint, aiming at enforcing desirable ambiguity function properties. Owing
to a suitable reformulation of the considered non-convex design problem, an
iterative optimization procedure over the transmit signal and the receive filter
has been devised.
Finally, the proposed method has been validated through experimental
results, showing the capability of the procedure to ensure the significant SINR
gains, with respect to the traditional approach adapting the receiver side only.
We have also analyzed the output SINR behavior against the TAA mismatch,
and the results have shown its robustness.
Adaptive Waveform Design of Polarimetric Radar 13
References
1. https://www.sdms.afrl.af.mil/index.php?collection=gtri
2. Aubry, A., Demaio, A., Farina, A., Wicks, M.: Knowledge-aided (potentially cog-
nitive) transmit signal and receive filter design in signal-dependent clutter. IEEE
Trans. Aerosp. Electron. Syst. 49(49), 93–117 (2013)
3. Chen, C.Y., Vaidyanathan, P.P.: Mimo radar waveform optimization with prior
information of the extended target and clutter. IEEE Trans. Sig. Process. 57(9),
3533–3544 (2009)
4. Garren, D.A., Odom, A.C., Osborn, M.K., Goldstein, J.S., Pillai, S.U., Guerci,
J.R.: Full-polarization matched-illumination for target detection and identification.
IEEE Trans. Aerosp. Electron. Syst. 38(3), 824–837 (2002)
5. Giuli, D.: Polarization diversity in radars. Proc. IEEE 74(2), 245–269 (1986)
6. Karbasi, S.M., Aubry, A., De Maio, A., Bastani, M.H.: Robust transmit code and
receive filter design for extended targets in clutter. IEEE Trans. Sig. Process. 63(8),
1965–1976 (2015)
7. Novak, L.M., Sechtin, M.B., Cardullo, M.J.: Studies of target detection algorithms
that use polarimetric radar data. IEEE Trans. Aerosp. Electron. Syst. 25(2), 150–
165 (1989)
8. Trees, H.L.V.: Detection of Signals-Estimation of Signal Parameters, vol. 4. John
Wiley & Sons Inc., New York (2002)
9. Ai, W., Huang, Y., Zhang, S.: Further results on rank-one matrix decomposition
and its application. Mathematical Programming, January 2008
10. Cheng, X., Aubry, A., Ciuonzo, D., De Maio, A., Wang, X.: Robust waveform and
filter bank design of polarimetric radar. IEEE Trans. Aerosp. Electr. Syst. 53(1),
370–384 (2017)
CRAWL: A Trace Routing Algorithm Based
on Hybrid Two-Layer Topology
1 Introduction
Data distribution has become an important part of distributed application, and its real
time must be improved due to the development of emergency response systems. For
example, in network centric operations [1], military information must be sent from
resource node to the interested target nodes as quickly as possible. The ShakeCast [2]
system of US Geological Survey (USGS) sends the damaged figures of public facilities
to users after the earthquake, and it is to help those victims choose the fastest escape
route.
A publish/subscribe model is a key technology for numerous data distribution
when emergency happens. According to different subscribe models, publish/subscribe
systems can be divided into four kinds based on channels, themes, types, and content
2 System Structure
To better take advantage of structured and unstructured topology, and offset their
weaknesses, CBDLO (Content-based Two-Layered Overlay) is presented to address
the problem of composing a content-based fast data distribution topology in the
dynamic network, and it integrates the characteristics of the flexibility and expand-
ability of unstructured topology, accuracy and low overhead on the information of
structured topology.
The lower layer of CBDLO is an unstructured net topology for the maintenance of net
connectivity, whose logic structure is similar to Random Graph. Every node keeps a
certain amount of other nodes as neighbors, which is addressed as View and update
regularly to enable the topology to timely reflect the net connectivity and dynamics.
In CBDLO, the way of random sampling of nodes and updating their views makes
topology preserve the structure of random graph and also maintain the characteristics of
small world model and power law. Meanwhile, every node is equipped with a property
topology APT (Access Point Lookup Table) and cache some information of upper
property topology to support the access of other nodes and the routing when transmit data.
16 L. Zheng et al.
CBDLO classifies nodes with the same subscribe property into clusters. Every
cluster is a property topology. According to the property value distribution, it also
divides the whole property value space into several connective but mutually exclusive
subspace. And every subspace corresponds to one virtual node, then all virtual nodes
are mapped to the real ones so as to make up of a distributed and balanced binary tree
structure.
In CBDLO, when data are generated, they firstly are forwarded through random
walk in the lower topology; after matching with the corresponding property topology,
they enter the upper property topology and begin to transmit data in a definite way.
Two-layered net topology structure is shown in Fig. 1.
upper layer:
property topology structure
Lower layer:
unstructured topology structure
In the lower-layer topology, data transmit by random walk and look for the matching
property topology on every node through the property topology APT. To avoid circuit
in the process of random walk, CRAWL (Content-based Routing Algorithm with
Labeled Trace) is put forward to record the latest routing by taking along a routing trace
buffer when transmit data. To avert path blocking brought by trace, routing recovery
mechanism is introduced.
node seeks for its corresponding property topology in APT. If it succeeds, data will be
directly forwarded to the corresponding property topology; if not, a neighbor node is
randomly chosen to transmit data (Fig. 2).
situation is illustrated in Fig. 3. In this figure, node A initiates a random walk along the
path following the arrows, B ! C ! D ! E ! F, but node F is on the edge of net
with a single neighbor node. Path return signifies that after the path has been blocked, a
node from Trace could be randomly chosen and then restart a random walk, but to
avoid being trapped into blocking and path recovery repeatedly, the maximum hop of
each new random walk should be reduced to half.
4 CDM Algorithm
When data find their upper property topology by means of random transmission in the
lower topology of CBDLO, data begin to accelerate matching distribution. To address
data matching distribution on the upper layer balanced binary tree topology, CDM
(Counter-based Distributed Matching Algorithm) is proposed. The matching data
properties on the upper layer distributed and balanced binary tree are identified through
CDM, then matching data could be sent to nodes interested in such data correspond-
ingly. Then subscribe nodes count the matching properties. If the number of matching
properties surpasses that number, data can be received.
Soon after data arrive at property topology, the only task is to find their matching
nodes, then transmit data. Property topology is a balanced binary tree structure, so its
matching process is actually fulfilled on binary tree, but before and after its trans-
mission on the binary tree, two conversions should be implemented.
The specific matching process is shown in Fig. 4. Step 1 represents that data nodes
should firstly confirm whether matching properties of the received data could be found
in the virtual node list; if they do, looking for virtual nodes including the property value
among virtual nodes is followed; if they do not, the subscriber probably has left, so a
new random walk on the lower topology should be restarted. In order to avoid being
stuck in such situation repeatedly, every start of new random walk should reduce the
maximum hop to half.
CRAWL: A Trace Routing Algorithm 19
property is a self-balancing binary tree, the height of the tree is equal to N, and N is the
total number of virtual nodes in the property topology.
5
x 10
4.5
4 Simple protocol
CRAWL algorithm
3.5
successful matching overhead
2.5
1.5
0.5
0
0 500 1000 1500 2000 2500 3000
total subscription number
Fig. 5. Effective distribution cost changing with the total number of subscribe
As shown in Fig. 5, the 1st point of Simple protocol demonstrates that the total
number of all messages in the system is 68605, and theoretically matched distributed
messages should be 408, while the actual matched messages are only 68. So the
overhead of effective distributive messages is 68605 / (68 / 408) = 411630.
In Fig. 5, CRAWL greatly reduces system’s effective distribution overhead.
Because in the random transmission process, the matching probability increases with
the rising amount of subscription, the effective distribution overhead of Simple protocol
gradually decreases with the increase of subscription. However, the increase of sub-
scription has rather slight influence on CBDLO, because in the simulation process
when the subscription reaches a certain amount, the message generated during a certain
time is still fixed. For CBDLO, CRAWL algorithm is affected only on the first stage.
The overhead on the second stage is fixed which equals the height of property topology
tree while the influence on the first stage is just on the number of property and CBDLO
CRAWL: A Trace Routing Algorithm 21
5
x 10
4.5
Simple
CRAWL
overhead of effective distributedive messages
3.5
2.5
1.5
0.5
10 15 20 25 30 35 40 45 50
max hops
Fig. 6. The relationship between overhead of effective distributed messages and the max hops
4
10
TERA
average overhead of distributed messages
CBDLO
3
10
2
10
1
10 2 3 4 5 6 7 8 9
10 10 10 10 10 10 10 10
nodes number
Figure 7 illustrates the effective distributive message overhead changes with nodes
by TERA and CBDLO respectively. In the experiments, nodes range from 100 to
1,000,000,000. TERA has 100 themes. According to the experiment of literature, the
random transmission hit rate of TERA lower topology goes up with the increase of
APT. When APT is 50, max hop of random walk is 10, the hit rate could be 90%. In the
large network, the hit rate will not improve obviously at the rise of APT, so APT of
TERA should be 50 taking space and efficiency into consideration.
22 L. Zheng et al.
The hit rate of CBDLO’s lower topology will go up with the rise of APT, but if the
hit rate will be 90% when the max hop is 10, then APT is no less than 100. So APT for
CBDLO is set as 100. Based on the above conditions, the experimental results are
shown in Fig. 7. It shows when network is small, with increasing nodes, the message
overhead difference of TERA and CBDLO is obvious, for which TERA’s broadcast
overhead is not so much in small network and CBDLO has many lower message
overhead; but with the increasing scale of network, TERA’s message overhead
increases quickly while CBDLO increases much slower, which shows the upper
structured topology of CBDLO effectively reduces message overhead. Thus, CBDLO
is better than TERA and more adaptable to the large scale network data development.
220
200
180
160
matching overhead
140
CBDLO/8
Simple/20
120
100
80
60
40
20
500 1000 1500 2000 2500 3000 3500 4000 4500 5000
subscription number
180
160
140
overhead of matching
Simple/8
120 CBDLO/20
100
80
60
40
20
60 70 80 90 100 110 120 130 140 150
number of properties
The changes of matching overhead with the property sum are shown in Fig. 9. The
vertical ordinate is dealt with the same way in Fig. 8. It remains to show the strong
advantages of CBDLO. The change of property number has a very slight influences on
CBDLO and this fully displays the effect of parallel matching.
6 Conclusion
Content-based fast data distribution is a key technology for constructing a large scale
distributive system and of great significance for data-centered application. However,
the construction of efficient and robust fast data distribution system confronts many
challenges, such as heterogeneity and dynamism of wide area network, diversity and
dynamism of data set, and the various customer requirements. Through the research of
fast data transmission in the dynamic network systems, a trace routing algorithm
CRAWL based on hybrid two-layer topology is proposed in this paper, which consists
of CBDLO. Besides, a distributed matching algorithm CDM (Counter-based Dis-
tributed Matching Algorithm) is also developed. The experimental results show that the
algorithm proposed in this paper applies to dynamic network with a strong adaptability,
high efficiency of avoiding circuit, good capability of reducing redundant message
overhead and matching delay, as well as improvement of distribution efficiency.
References
1. Dell-Imagine, R.: JTIDS-an overview of the system design and implementation (Joint
Tactical Information Distribution System). In: Position Location and Navigation Sympo-
sium, San Diego, California, pp. 212–215 (2013)
2. ShakeCast [EB/OL] (2014). http://earthquake.usgs.gov/research/software/-shakecast/
3. Eugster, P.T., Felber, P., Guerraoui, R., et al.: The many faces of publish/subscribe. ACM
Comput. Surv. (CSUR) 35(2), 114–131 (2003)
4. Castro, M., Druschel, P., Kermarrec, A., et al.: SCRIBE: a large-scale and decentralized
application-level multicast infrastructure. IEEE J. Sel. Areas Commun. 20(8), 1489–1499
(2012)
24 L. Zheng et al.
5. Rahimian, F., Girdzijauskas, S., Payberah, A.H., et al.: Vitis: a gossip-based hybrid overlay
for internet-scale publish/subscribe enabling rendezvous routing in unstructured overlay
networks. In: IEEE International Parallel & Distributed Processing Symposium (IPDPS),
pp. 746–757 (2011)
6. Zheng, Z., Wang, Y., Ma, X.: PeerChatter: a peer-to-peer architecture for data distribution
over social networks. Inf.-Int. Interdisc. J. 15(1), 259–266 (2011)
7. Pietzuch, P., Hermes, B.J.: A distributed event-based middleware architecture. In: 22nd
International Conference on Distributed Computing Systems Workshops, pp. 611–618
(2012)
8. Gupta, A., Sahin, O.D., Agrawal, D., et al.: Meghdoot: content-based publish/subscribe over
P2P networks. In: Proceedings of the 5th ACM/IFIP/USENIX international conference on
Middleware (Middleware), pp. 254–273 (2004)
9. Grummt, E.: Fine-grained parallel XML filtering for content-based publish/subscribe
systems. In: Proceedings of the 5th ACM International Conference on Distributed Event-
Based System, pp. 219–228 (2011)
10. Li, M., Ye, F., Kim, M., et al.: A scalable and elastic publish/subscribe service. In: IEEE
International Parallel & Distributed Processing Symposium (IPDPS), pp. 1254–1265 (2011)
11. Cao, F., Singh, J.P.: MEDYM: match-early with dynamic multicast for content-based
publish-subscribe networks. In: Proceedings of the ACM/IFIP/USENIX International
Conference on Middleware (Middleware), pp. 292–313 (2005)
12. Wong, B., Guha, S.: Quasar: a probabilistic publish-subscribe system for social networks. In:
Proceedings of the 7th International Workshop on Peer-to-Peer Systems, pp. 2–7 (2008)
13. Aekaterinidis, I., Triantafillou, P.: PastryStrings: a comprehensive content-based
publish/subscribe DHT network. In: IEEE 26nd International Conference on Distributed
Computing Systems (ICDCS), pp. 23–32 (2006)
14. Baldoni, R., Marchetti, C., Virgillito, A., et al.: Content-based publish-subscribe over
structured overlay networks. In: IEEE 25nd International Conference on Distributed
Computing Systems (ICDCS), pp. 437–446 (2014)
Design and Implementation of IoT
Based Class Attendance Monitoring System
Using Computer Vision and Embedded
Linux Platform
1 Introduction
The Automated Attendance System is the progress that has taken place in the field of
automation replacing the attendance marking activity traditionally. The traditional
method of attendance marking is very time consuming and becomes difficult when the
strength is high. Automation of attendance system has boundary over the conventional
approach as it saves time and also can be used for security purposes. This system also
helps to prevent fake attendance. So automating the attendance process will increase
the productivity of the class and make good attendance records.
2 Related Work
In this section, we reviewed some related technologies, innovative and previous works
on the topic of attendance monitoring system based application.
The most widely used location-based technologies are GPS, Radio Frequency
Identification (RFID), Bluetooth and so on. Researchers work on these technologies for
improving the services for its accuracy, reliability, and efficiency [1]. A lot of appli-
cation exists in different ways to adequately monitor the attendance of every student.
Usually, attendance management uses student ID for authentication. Many researchers
[2] have designed and implemented an ID-based authentication system. This kind of
system required improvement as IDs can be shared or tampered.
Other attendance monitoring techniques use face detection based student atten-
dance system. Problem with face detection based systems is that individuals have to
carry face detection which is not a cost-effective solution [3]. However, the
improvement of the system has been discovered by M. Wong. By introducing palm-
print which developed an attendance system to record employee attendance [4].
Meanwhile, a similar project has been implemented which applied wireless Iris
recognition attendance management system. Biometric identification security is the
primary concern as if one lost its biometric identification from the database then the
user of the system has to face many challenges. Mohammad et al. have defined location
based time and attendance system, that use location as the key of attendance [5]. The
employees in an organization can be determined with the help of a GPS device if both
Design and Implementation of IoT Based Class Attendance Monitoring System 27
coordinates match then the employee is present in the organization. Which can be
extended for different applications and on all mobile phones.
3 System Architecture
The architecture of our class attendance system based on the IoT is shown in Fig. 1.
The whole system is implemented on the embedded Linux platform using Rasp-
berry Pi board. Raspberry Pi is used as core processor, complete final processing of
images, data storage, real-time face detection and recognition and data server, after that
the data will be present on the network eventually. The power supply is given to the
Raspberry Pi which is the central part of the proposed system. The Pi camera is
connected to the Raspberry Pi CSI port. The camera captures the images of the students
who are present in the class. Raspberry Pi takes those images as input images and
compares the input images with the existing image dataset. This process happens due to
importing the open CV packages at the initial stage of the development of the system.
After detection and recognition, the system generates an excel sheet containing student
attendance result. All of these processes can be controlled by using a web application.
The system collects Excel file and uploads the data into the cloud server which has
been developed by flask web framework in Raspberry Pi. Real-time data visualization
can be accessed and downloaded directly from the cloud server through the web
application.
As a part of the Internet of Things, it requires network connectivity through WiFi
which is connected with Raspberry Pi. The web application can be used by the internet
and collect attendance result with minimum human intervention. The system is also
capable of sending the result by email to respective teacher and student.
28 H. Salman et al.
4 Proposed Approach
The proffered system management will use three step mechanisms: Dataset creation,
training dataset, testing.
In dataset creation, the system Initialize the camera connected to Raspberry Pi, get
image and user id as input. After converting the image into grayscale, it detects the
face’s images of the students in a specific class, who are present in the classroom and
store it in the dataset.
In training dataset part it initializes LBPH face recognizer and gets faces and ID’s
from dataset folder to train the LBPH face recognizer. After training data, it is stored as
XML or YML file. In testing part the system load Haar classifier, LBPH face recog-
nizer and trained data from XML or YML file. Then capturing the image from a
camera, it converts it into grayscale, detects and predict the face using the above
recognizer. If the system recognizes a student’s face, it will be reported or recorded as
present and all those not in the class or not recognize will be absent.
Then, the system collects these attendance data and generate an Excel file and
upload the file in the cloud server. Finally, it sends email to teachers and students
containing attendance data from the process. The flow chart of the proposed approach
is given in Fig. 2.
The Raspberry Pi is the central part of the project. All the data and algorithm has
been processed and programmed with python language in Raspberry Pi. The Pi camera
has been used to capture photos. A cloud server has been developed by using the flask
web framework of python. To access the class attendance system from smartphone or
PC, a user-friendly web application has been developed using HTML and
Design and Implementation of IoT Based Class Attendance Monitoring System 29
Bootstrap. The main aim of this proffered management system: giving a vivid and
accurate attendance record of the students, it prevents human intervention concerning
the attendance record, mailing students and teachers using the easiest way of checking
attendance of students in a specific classroom.
5 Hardware Requirements
5.1 Raspberry Pi 3 Model B
Raspberry Pi is an ARM-based single board computer. The Raspberry Pi 3 Model B is
used in this paper, which is the third generation Raspberry Pi [6]. The operating system
used for Raspberry Pi is Raspbian as it is open source. Raspbian is a Linux-based
computer operating system. It has 40 pins in which 24 are GPIO pins and these pins are
used for general purpose.
5.2 Pi Camera
Pi Camera is a custom designed add-on for Raspberry Pi. This interface uses the CSI
interface designed especially for interfacing to cameras [7]. The sensor has 5
megapixels of resolution and has a fixed focus lens on board.
6.1 OpenCV
OpenCV is a computer vision software library. The library has a lot of optimized
algorithms, which can be used in many IoT related sectors including face detection and
recognition. As the libraries of our project, we use the Haar classifier and LBPH face
recognizer.
the concept of integral image has been used. The integral image is derived by using the
following equation:
IR ðx; yÞ ¼ R iðx0 ; y0 Þ
x0 x
y0 y
6.3 Python
Python is a high-level programming language which is an open source working for a
general purpose language. It can be embedded into any application and can run on all
current operating systems. Overall Python is suitable as a development language for
customizable applications.
6.4 Numpy
Numpy is a library for scientific computing in Python. It provides a high-performance
multidimensional array object and matrices, and along with an extensive collection of
tools to operate on these arrays. For converting images of our project into multidi-
mensional or 2D-array representation, and conversions from grayscale to color images.
7 Experimental Result
The IoT based class attendance monitoring system using an embedded Linux platform
functionally implemented, and results obtained successfully.
Fig. 5. Result of face detection (Data set) and face recognition (Right side).
32 H. Salman et al.
Fig. 6. Class attendance system based on IoT (Left side) and python application (Right side).
Fig. 7. Web application for user interface and result of attendance system.
Figure 5 shows the system create a dataset for face detection and recognition and
the result of multiple face recognition. In Table 1, the system accuracies data are given.
The data has been collected and verified among 30 students. Figure 6 shows com-
pletion of smart attendance system based on IoT and python application. Figure 7
shows the web application for user interface and result of attendance in excel sheet
which can be downloaded from the application using Internet from smartphone or
Design and Implementation of IoT Based Class Attendance Monitoring System 33
computer. The user can register and check attendance for smart class attendance by
clicking buttons of the web application. The system uses SMTP protocol for sending
email to the students and teacher in a specific time. All of these features make the IoT
based system efficient enough for taking attendance of any class. This approach is done
so that in future as the number of students increases the system does not fails and can
provide backup and remain accurate and reliable in the long run. Also, to use the most
efficient algorithm for the system with the use of upcoming enhanced algorithms in
future, the accuracy of the system can be enhanced.
8 Conclusion
IoT based class attendance monitoring system is a time-saving, reliable and useful
technique for taking attendance instead of the manual technique. There are many
systems which can be used for managing attendance, but the implementation of face
detection and recognition system with the IoT and Raspberry Pi technology make the
system economy. Since the proposed system gets integrated with the cloud server, it
becomes system independent, fast, provides data storage and accessed quickly. The
web application and email notification make the system user-friendly. In the future, we
will improve the attendance record system according to the following:
(1) There will be an automatic printing of student’s attendance record for the teaching
staff.
(2) To monitor students during an examination.
(3) To trace a particular student in an organization quickly with the help of this
system and can be utilized for several security applications where authentication is
required to access the privileges of the respective system.
References
1. Patel, R., Patel, N., Gajjar, M.: Online students attendance monitoring system in classroom
using radio frequency identification technology: a proposed system framework. Int.
J. Emerg. Technol. Adv. Eng. 2(2), 61–66 (2012)
2. Gowri, Ch.S.R., Kiran, V., Rama Krishna, G.: Automated intelligence system for attendance
monitoring with open CV based on internet of things (IoT). Int. J. Sci. Eng. Technol. Res.
(IJSETR) 5(4), 905–913 (2016)
3. Shoewu, O., Olaniyi, O.M., Lawson, A.: Embedded computer-Based lecture attendance
management system. Afr. J. Comput. ICT 4(3), 27–36 (2011)
4. Mani Kumar, B., Praveen Kumar, M., Rangareddy: RFID based Attendance monitoring
system using IOT with TI CC3200 Launchpad. Int. J. Mag. Eng. Technol. Manag. Res. 2(7),
1465–1467 (2015)
34 H. Salman et al.
5. Uddin, M.S., Allayear, S.M., Das, N.C., Talukder, F.A.: A location based time and
attendance system. Int. J. Comput. Theor. Eng 6(1), 1–2 (2014)
6. Grimmett, R.; Raspberry Pi Robotic Projects. 3rd Edn. Packt Publishing (2016)
7. Abaya, W.F., Basa, J., Sy, M., Abad, A.C., Dadios, E.P.: Low cost smart security camera
with night vision capability using Raspberry Pi and OpenCV. In: 2014 International
Conference on Humanoid, Nanotechnology, Information Technology, Communication and
Control, Environment and Management (HNICEM), Palawan, pp. 1–6 (2014)
8. Pasumarti, P., Purna Sekhar, P.: Classroom attendance using face detection and Raspberry-
Pi. Int. Res. J. Eng. Technol. (IRJET) 05(03), 3–5 (2018)
9. Rajkumar, S., Prakash, J.: Automated attendance using Raspberry Pi. Int. J. Pharm. Technol.
(IJPT) 8(3), 16214–16221 (2016)
10. T. M. Inc.: Train a Cascade Object Detector. http://www.mathworks.se/help/vision/ug/train-
a-cascadeobject-detector.html#btugex8
11. Ronacher, A.: Quickstart (2010). http://flask.pocoo.org/docs/0.12/quickstart/#quickstart
12. http://getbootstrap.com/
An Indoor 3D Positioning Technology
Based on NB-IoT
Abstract. The Narrow Band Internet of Things (NB-IoT) for low-power, wide-
coverage, and low-cost requirements solves the problem of massive Internet of
Things device connectivity. NB-IoT supports device connection with large data
throughput and long standby time. It is unmatched by traditional cellular data
network technology and Bluetooth technology. By combining the NB-IoT
positioning with the Lagrangian multiplier improved constrained least squares
localization algorithm and the GSM mobile terminal delay estimation algorithm
in NLOS environment, a feasible indoor three-dimensional positioning method
is proposed. In addition, it reduces the influence of environmental factors on
positioning and improves the accuracy of indoor three-dimensional positioning.
1 Introduction
Based on the concepts of smart city, big data, and cloud computing, the Internet of
Things technology has developed rapidly. Various physical devices are connected to
the Internet to transmit large amounts of data over the Internet to enable monitoring and
management of physical devices. Massive connections of various types of physical
equipment, and the data transmitted are also diverse. The requirements for transmission
have also increased, and various types of transmission equipment have begun to
integrate. Different from traditional cellular data network technology, Bluetooth tech-
nology [1] and other remote transmission technologies. Under such a drive, Narrow
Band Internet of Things (NB-IoT) [2] was released by the 3GPP, which is the wireless
technology standardization organization. As an evolutionary technology of LTE,
NB-IoT not only has a higher transmission rate, but also has a large number of device
connectivity capabilities. Features such as lower latency and lower power consumption
are the top choices for manufacturers, and this accelerates the application of the Internet
of Things in life.
The indoor three-dimensional positioning, as the final need to overcome the
positioning technology, has a very broad application prospects. For example: indoor
navigation, garage positioning, etc. However, when the device is indoors, the receiving
satellite signal is easily interfered by the obstruction and causes a non-line-of-sight
error, which makes it impossible to accurately obtain the positioning data [3]. In
addition, indoor complex electromagnetic interference has become one of the important
factors affecting positioning. Various factors have limited the further development of
current indoor positioning technology. The emergence of NB-IoT solves the problem.
It utilizes the characteristics of NB-IoT, large connection and deep coverage to solve
the problem of indoor receiving satellite signal receiving interference. At the same
time, NB-IoT has low cost and low power consumption. The characteristics make
NB-IoT more extensible and can be widely used in the current Internet of Things
industry.
In order to improve the fusion of technology and improve the indoor positioning
accuracy, this paper uses the Lagrangian multiplier improved constrained least squares
positioning algorithm in the newly proposed NLOS environment to solve the non-line-
of-sight error problem. And the delay estimation algorithm of the GSM mobile station
reduces the dependence of the terminal equipment on the base station and implements
passive positioning with low dependency.
3 Location Algorithm
Where di is the moving distance between the reference node and the target, which
ðxi ; yi Þ is the coordinate of the current position of the target, and i is a constant. Then
square the two sides of the equation to get the current actual distance.
The equal sign is established only in an ideal environment where there is no line-of-
sight error. Therefore, in the real environment, only the greater than the relationship
is taken.
Then, based on the values of di obtained in the Eq. (1), the matrix A is constructed,
and the estimated value obtained of v from the result of A is obtained.
42 D. Xue et al.
1
LðX; aÞ ¼ X T HX þ CT X aT ðBXÞ ð4Þ
2
According to the Eq. (4), the value of a is brought into the Eq. (5), and the CLS
non-line-of-sight error elimination algorithm based on the Lagrangian multiplier can be
obtained.
P1
Z2 ¼ Z1 þ 18400 ð1 þ a tÞ lg ð6Þ
P2
In the above formula, P2 is the pressure at the current position of the target, P1 is
the ground standard reference air pressure, Z1 is the ground height, the initial value is
0, Z2 is the height of the current position of the target, and a and t are constant.
When the target moves upstairs or downstairs, the air pressure at the location is
continuously updated, and then the height is automatically converted according to the
pressure height formula to monitor the current position of the target in real time.
Because time, weather and other factors will also affect the target’s air pressure at the
same position, but the pressure value changes within the acceptable error range, and
after the high pressure formula conversion, the error will be further reduced, so this
article is only for the target The data conversion estimation is performed at the position
when the height conversion is performed.
An Indoor 3D Positioning Technology Based on NB-IoT 43
4 Conclusion
Aiming at the problems of non-line-of-sight error and multipath effect and delay
estimation caused by signal quality caused by measurement noise and obstruction
during transmission in indoor three-dimensional positioning, the Lager in the NLOS
environment is proposed based on NB-IoT. The Lange multiplier improves the least
squares algorithm and the mobile station delay estimation location algorithm. When
solving indoor positioning and receiving signals, it is affected by interference factors
and cannot be accurately positioned. At the same time, it can effectively and accurately
realize the three-dimensional positioning of measuring the target height in indoor
environment.
References
1. Hallberg, J., Nilsson, M., Synnes, K.: Positioning with bluetooth. In: 10th IEEE International
Conference on Telecommunications, pp. 954–958 (2003)
2. Huang, W.: Current status and development trend of NB-IoT low-rate narrowband internet
of things communication technology. Electron. Test (6), 58–29 (2017)
3. Mahmud, M.S., Qaisar, S.U., Benson, C.: Weak GPS signal detection in the presence of
strong signals with varying ralative Doppler and long integration gain. In: IEEE/ION
Position, Location and Navigation Symposium, pp. 1015–1020 (2016)
4. Liu, Y., Yang, Z., Wang, X., Jian, L.: Location, localization, and localizibility. J. Comput.
Sci. Technol. 25(2), 274–297 (2010)
5. Gu, Y., Chen, Y., Liu, J., Jiang, X.: Semisupervised deep extreme learning machine for
Wi-FI based localization. Neurocomputing 166, 282–293 (2015)
6. Wu, X., Shen, S., Zhao, J.: Research on wireless sensor network convergence network based
on NB-IoT technology. Comput. Appl. Softw. 35(6), pp. 1000–386 (2018)
7. Dai, B., Yuan, Y., Yu, Y.: Narrowband Internet of Things (NB-IoT) Standards and Key
Technologies. People’s Posts and Telecommunications Press, Beijing (2016)
8. Zhai, X., Wu, C.: Time delay estimation algorithm for time difference positioning in GSM
mobile stations. J. Comput. Appl. 36, 22–24 (2016)
9. Li, S., Hua, J., Wang, D., Zhou, K., Chen, F., Yu, X.: Improved Constrained least squares
localization algorithm for lagrangian multiplier in NLOS environment. Chin. J. Sens.
Actuators 31(8), 1004–1699 (2018)
10. Zheng, Z., Hua, J., Jiang, B., et al.: A new non-line-of-sight suppression wireless localization
algorithm using the optimization principle. Chin. J. Sens. Actuators 26(5), 722–727 (2013)
Consideration of Implementation
Method for the Cloud Type Virtual Policy
Based Network Management
Scheme for the Specific Domain
Abstract. In the current Internet system, there are many problems using
anonymity of the network communication such as personal information leaks
and crimes using the Internet system. This is why TCP/IP protocol used in
Internet system does not have the user identification information on the com-
munication data, and it is difficult to supervise the user performing the above
acts immediately. As a study for solving the above problem, there is the study of
Policy Based Network Management (PBNM). This is the scheme for managing
a whole Local Area Network (LAN) through communication control for every
user. In this PBNM, two types of schemes exist. As one scheme, we have
studied theoretically about the Destination Addressing Control System (DACS)
Scheme with affinity with existing internet. By applying this DACS Scheme to
Internet system management, we will realize the policy-based Internet system
management. In this paper, to realize management of the specific domain with
some network groups with plural organizations, concept of implementation
method applied for this scheme is described.
1 Introduction
In the current Internet system, there are many problems using anonymity of the net-
work communication such as personal information leaks and crimes using the Internet
system.
As a study for solving the problems, Policy Based Network Management (PBNM)
[2] exists. The PBNM is a scheme for managing a whole Local Area Network
(LAN) through communication control every user, and cannot be applied to the Internet
system. In the existing PBNM, there are two types of schemes. The first is the scheme
As a study for solving the problem on Internet System, the study area about PBNM
exists. This is a scheme of managing a whole LAN through communication control
every user. Because this PBNM manages a whole LAN by making anonymous com-
munication non-anonymous, it becomes possible to identify the user who steals per-
sonal information and commits a crime swiftly and easily. Therefore, by applying this
policy-based thinking, we study about the policy-based Internet system management.
In policy-based network management, there are two types of schemes. The first
scheme is the scheme described in Fig. 1. The standardization of this scheme is per-
formed in various organizations. In IETF, a framework of PBNM [2] was established.
Standards about each element constituting this framework are as follows. As a model of
control information stored in the server called Policy Repository, Policy Core Infor-
mation model (PCIM) [3] was established. After it, PCMIe [4] was established by
46 K. Odagiri et al.
extending the PCIM. To describe them in the form of Lightweight Directory Access
Protocol (LDAP), Policy Core LDAP Schema (PCLS) [5] was established. As a pro-
tocol to distribute the control information stored in Policy Repository or decision result
from the PDP to the PEP, Common Open Policy Service (COPS) [6] was established.
Based on the difference in distribution method, COPS usage for RSVP (COPS-RSVP)
[7] and COPS usage for Provisioning (COPS-PR) [8] were established. RSVP is an
abbreviation for Resource Reservation Protocol. The COPS-RSVP is the method as
follows. After the PEP having detected the communication from a user or a client
application, the PDP makes a judgmental decision for it. The decision is sent and
applied to the PEP, and the PEP adds the control to it. The COPS-PR is the method of
distributing the control information or decision result to the PEP before accepting the
communication.
Next, in DMTF, a framework of PBNM called Directory-enabled Network
(DEN) was established. Like the IETF framework, control information is stored in the
server storing control information called Policy Server, which is built by using the
directory service such as LDAP [9], and is distributed to network servers and net-
working equipment such as switch and router. As the result, the whole LAN is man-
aged. The model of control information used in DEN is called Common Information
Model (CIM), the schema of the CIM (CIM Schema Version 2.30.0) [11] was opened.
The CIM was extended to support the DEN [10], and was incorporated in the
framework of DEN.
In addition, Resource and Admission Control Subsystem (RACS) [12] was
established in Telecoms and Internet converged Services and protocols for Advanced
Network (TISPAN) of European Telecommunications Standards Institute (ETSI), and
Resource and Admission Control Functions (RACF) was established in International
Telecommunication Union Telecommunication Standardization Sector (ITU-T) [13].
However, all the frameworks explained above are based on the principle shown in
Fig. 1. As problems of these frameworks, two points are presented as follows. Essential
principle is described in Fig. 2. To be concrete, in the point called PDP (Policy
Decision Point), judgment such as permission and non-permission for communication
Consideration of Implementation Method for the Cloud Type Virtual Policy 47
pass is performed based on policy information. The judgment is notified and trans-
mitted to the point called the PEP, which is the mechanism such as VPN mechanism,
router and Fire Wall located on the network path among hosts such as servers and
clients. Based on that judgment, the control is added for the communication that is
going to pass by.
The principle of the second scheme is described in Fig. 3. By locating the com-
munication control mechanisms on the clients, the whole LAN is managed. Because
this scheme controls the network communications on each client, the processing load is
low. However, because the communication control mechanisms need to be located on
each client, the work load becomes heavy.
When it is thought that Internet system is managed by using these two schemes, it is
difficult to apply the first scheme to Internet system management practically. This is why
the communication control mechanism needs to be located on the path between network
servers and clients without exception. On the other hand, the second scheme locates the
communication controls mechanisms on each client. That is, the software for commu-
nication control is installed on each client. So, by devising the installing mechanism
letting users install software to the client easily, it becomes possible to apply the second
scheme to Internet system management. As a first step for the last goal, we showed the
Wide Area DACS system (wDACS) system [15]. This system manages a wide area
network, which one organization manages. Therefore, it is impossible for plural orga-
nizations to use this system. Then, as the next step, we showed the cloud type virtual
PBNM, which could be used by plural organizations in this paper.
According to the distributed DACS rules, the DACS Client performs (1) or
(2) operation as shown in the following. Then, communication control of the client is
performed for every login user.
(1) Destination information on IP Packet, which is sent from application program, is
changed.
(2) IP Packet from the client, which is sent from the application program to the
outside of the client, is blocked.
An example of the case (1) is shown in Fig. 4. In Fig. 4, the system administrator
can distribute a communication of the login user to the specified server among servers
A, B or C. Moreover, the case (2) is described. For example, when the system
administrator wants to forbid an user to use MUA (Mail User Agent), it will be
performed by blocking IP Packet with the specific destination information.
In order to realize the DACS Scheme, the operation is done by a DACS Protocol as
shown in Fig. 5. As shown by (1) in Fig. 5, the distribution of the DACS rules is
performed on communication between the DACS Server and the DACS Client, which
is arranged at the application layer. The application of the DACS rules to the DACS
Control is shown by (2) in Fig. 5.
3.4 The Cloud Type Virtual PBNM for the Common Use Between Plural
Organizations
In this section, after the concept and implementation of the proposed scheme were
described, functional evaluation results are described. In Fig. 8 which is described in
[16], the proposed concept is shown. Because the existing wDACS Scheme realized the
PBNM control with the software called the DACS Server and the DACS client, other
mechanism was not needed. By this point, application to the cloud environment was
easy.
Fig. 8. Cloud type virtual PBNM for the common use between plural organizations
The proposed scheme in this paper realizes the common usage by plural organi-
zations by adding the following elements to realize the common usage by plural
organizations: user identification of the plural organizations, management of the policy
information of the plural organizations, application of the PKI for code communication
in the Internet, Redundant configuration of the DACS Server (policy information
server), load balancing configuration of the DACS Server, installation function of
DACS Client by way of the Internet In the past study [14], the DACS Client was
operated on the windows operation system (Windows OS). It was because there were
many cases that the Windows OS was used for as the OS of the client. However, the
Linux operating system (Linux OS) had enough functions to be used as the client
recently, too. In addition, it was thought that the case used in the clients in the future
came out recently. Therefore, to prove the possibility of the DACS Scheme on the
Linux OS, the basic function of the DACS Client was implemented in this study. The
basic functions of them were implemented by JAVA language.
Consideration of Implementation Method for the Cloud Type Virtual Policy 51
ApplicaƟon of ApplicaƟon of
the policy informaƟon the policy informaƟon
Network of the own Network of the own Network of the own Network of the own
organizaƟon organizaƟon organizaƟon organizaƟon
(Org. A) (Org. B) (Org. C) (Org. D)
Domain 1
Movement terminal in Org. A
(1)Movement and connecƟon
(2)Usage depending on policy informaƟon
The user authentication referred to here is the one before distribution of DACS
rules as policy information. It is explicitly necessary to distinguish it from user
authentication for permitting network connection as a separate one. Depending on the
implementation, it may be descriptively possible to integrate these two user authenti-
cations as one process, but this is a matter to be considered at the implementation stage.
As Factor2, concept of policy information decision method is described. In Fig. 11,
as an example, the user belonging to the organization A connects the notebook client
computer to another organization in Network Group 2. In this case, the DACS rules as
policy information are extracted from the DACS SV as policy information server which
Consideration of Implementation Method for the Cloud Type Virtual Policy 53
Policy Policy
AuthenƟcaƟon InformaƟon AuthenƟcaƟon
InformaƟon RelaƟon
Server Server Server
Server
Domain 1 (b)
ApplicaƟon of
Policy
InformaƟon
User in Org. A User in Org. A
are located in both network group (Network Group 1 and Network Group 2). After it,
the DACS rules are determined and applied for the DACS CL in the client computer.
When the DACS rules are determined, multiple points are considered as follows.
(Point1) Use of the information service of the organization to which the user
belongs
(Point2) Use of the information service of the organization to which the user
connects
(Point3) Use of the Packet Filtering Service
As the first example, with the user in organization A connected the own client to the
network of network group 2, the user accesses the mail system in own organization. At
that time, to enable the client to access the mail service, the communication must be
permitted by packet filtering setting. In the setting of the mail client in the client, the
server name of the mail system of its own organization is normally specified. As the
result, it becomes possible to access the mail system from the client. In this case, as the
packet filtering setting, the policy information transmitted from the policy information
server of the management organization X is reflected on the client. As the second
example, with the user in organization A connected the own client to the network of
network group 2, the user accesses the mail system of the connected organization, like
the first example, to enable the client to access the mail service, the communication
must be permitted by packet filtering setting. In addition, because the server name of
the mail system of its own organization is normally specified in the setting of the mail
client in the client, it is necessary to change the mail system to be accessed by com-
munication control by NAT. As the policy information for these two communication
controls, the one transmitted from the policy information server of the connected
organization is reflected to the client. After it, it becomes possible to access the mail
system of the connected organization from the client. However, in the case of example
2, the policy information used in example 1 also exists, the policy information used in
54 K. Odagiri et al.
the above two examples is set in both management organizations. In this case, the fact
that the policy information is set in the connected organization means that the service
use of the connected organization is permitted.
the DACS Client. When the control information in them overlap, one takes precedence.
In principle, control rules that exist in the DACS Server on the network of the orga-
nization to which the user actually connects are prioritized. However, based on the
judgment of the organization, it becomes also possible to prioritize the control rules of
the organization to which the user belongs. For example, the following two control
rules are prepared. Rule 1 is control rule in the DACS Server in the organization to
which the user normally belongs. Rule 2 is control information in the DACS Server in
the organization to which the user is actually connected.
(Rule 1) Destination IP address and port before conversion: 133.20.25.10:80
Destination IP address and port after conversion: 133.20.25.11:80
(Rule 2) Destination IP address and port before conversion: 133.20.25.10:80
Destination IP address and port after conversion: 133.20.25.12:80
According to the above principle, since the control information of Rule 2 takes
precedence, this information is applied to the DACS Client. As a result, when
launching the Web browser and accessing the 133.20.25.10: 80 Web page, actually, the
Web page of 133.20.25.12: 80 is accessed. However, when setting information indi-
cating that rule 1 is prioritized, rule 1 takes precedence and is applied to the DACS
Client.
5 Conclusion
In this paper, implementation Method for the Cloud Type Virtual Policy Based Net-
work Management Scheme for the Specific Domain is described. Considering affinity
with the Internet system, the distributed authentication method was examined. In the
future, implementation will be carried out with details of the authentication method
stuffed up. In the future, implementation will be carried out based on this paper.
Acknowledgments. This work was supported by the research grant of JSPS KAKENHI Grant
Number 17K00113, and the research grant by Support Center for Advanced Telecommunications
Technology Research, Foundation (SCAT). We express our gratitude.
References
1. Cerf, V., Kahn, E.: A protocol for packet network interconnection. IEEE Trans. Commn.
COM-22, 637–648 (1974)
2. Yavatkar, R., Pendarakis, D., Guerin, R.: A Framework for Policy-based Admission Control.
IETF RFC 2753 (2000)
3. Moore, B., et al.: Policy Core Information Model – Version 1 Specification. IETF RFC 3060
(2001)
4. Moore, B.: Policy Core Information Model (PCIM) Extensions. IETF 3460 (2003)
5. Strassner, J., Moore, B., Moats, R., Ellesson, E.: Policy Core Lightweight Directory Access
Protocol (LDAP) Schema. IETF RFC 3703 (2004)
6. Durham, D., et al.: The COPS (Common Open Policy Service) Protocol. IETF RFC 2748
(2000)
56 K. Odagiri et al.
7. Herzog, S., et al.: COPS usage for RSVP. IETF RFC 2749 (2000)
8. Chan, K., et al.: COPS Usage for Policy Provisioning (COPS-PR). IETF RFC 3084 (2001)
9. CIM Core Model V2.5 LDAP Mapping Specification (2002)
10. Wahl, M., Howes, T., Kille, S.: Lightweight directory access protocol (v3). IETF RFC 2251
(1997)
11. CIM Schema: Version 2.30.0, 2011
12. ETSI ES 282 003: Telecoms and Internet converged Services and protocols for Advanced
Network (TISPAN); Resource and Admission Control Subsystem (RACS); Functional
Architecture, June 2006
13. ETSI ETSI ES 283 026: Telecommunications and Internet Converged Services and
Protocols for Advanced Networking (TISPAN); Resource and Admission Control; Protocol
for QoS reservation information exchange between the Service Policy Decision Function
(SPDF) and the Access-Resource and Admission Control Function (A-RACF) in the
Resource and Protocol specification, April 2006
14. Odagiri, K., Yaegashi, R., Tadauchi, M., Ishii, N.: Secure DACS scheme. J. Network
Comput. Appl. 31(4), 851–861 (2008)
15. Odagiri, K., Shimizu, S., Takizawa, M., Ishii, N.: Theoretical suggestion of policy-based
wide area network management system (wDACS system part-I). Int. J. Networked Distrib.
Comput. (IJNDC) 1(4), 260–269 (2013)
16. Odagiri, K., Shimizu, S., Ishii, N., Takizawa, M.: Suggestion of the cloud type virtual policy
based network management scheme for the common use between plural organizations. In:
Proceedings of International Conference on International Conference on Network-Based
Information Systems (NBiS-2015), pp. 180–186, September 2015
A Model of Virtual Machines to Support
Storage Processes
1 Introduction
Applications on clients use computation resources supported by a cluster of
servers like cloud computing systems [14]. A cluster of servers provides applica-
tions with virtual computation services through virtual machines [3]. That is,
applications can take usage of computation resources like CPUs and storages
like HDD through virtual machines without being conscious of heterogeneity
and location of each server in a cluster. In addition, virtual machines migrate
from a host server to a guest server in the live manner [3]. That is, application
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 57–66, 2019.
https://doi.org/10.1007/978-3-030-15035-8_6
58 N. Ogura et al.
2 System Model
2.1 Virtual Machines
2.2 Clones
A clone vml of a virtual machine vmk can be created on the host server sh of
the virtual machine vmk . Here, the virtual machine vmk is a parent of the clone
vml . There are two types of clones, full and linked clones [2]. If a full clone vml
is created from the parent virtual machine vmk , not only the virtual memory
mk but also files in the virtual storage vsk of the virtual machine vmk are copied
to the full clone vml . Full clones of a virtual machine are independent of one
another.
A linked clone vml shares virtual storage with a snapshot of a parent virtual
machine vmk . When the clone vml is created, the virtual storage of the clone
vml is a snapshot of vml . Files updated by the parent vmk do not affect the
clone vml and files updated by the clone vml do not affect the parent vmk ,
either. It takes shorter time to create a linked clone than a full clone.
In this paper, we consider a full clone.
3 Experiments
Table 1. Servers.
Servers sh sg
CPU Intel Core i5-8400 Intel Core i5-8400
Memory [GB] 8 8
HDD [TB] 1.0 1.0
OS CentOS 7.4 CentOS 7.4
A Model of Virtual Machines to Support Storage Processes 61
Fig. 2. Servers.
network (LAN) as shown in Fig. 2. The same operating system Linux CentOS
7.4 [1] is installed in a pair of the servers sh and sg .
A virtual machine vmk is realized in KVM [3] and supports CentOS 7.4 [1].
The server sh is a home server of the virtual machine vmk . The virtual machine
vmk provides the virtual memory mk and the virtual storage vsk . A virtual
machine vmk migrates from a host server sh to a guest server sg by issuing a
command “virsh migrate” on the host server sh . In addition, a clone vml of a
virtual machine vmk can be created on the host server sh of vmk .
We consider a storage process pi which reads data in a file fi as shown in
Fig. 3. The file fi is created by an application process on a virtual machine vmk
in the server sh . The virtual machine vmk is a provider of the file fi . In the
process pi , the file fi is first opened by an open system call. Here, the size
of the file fi is 5, 15, 20 [GB]. Then, data in the file fi is iteratively read by
read system calls. Data of 1,024[B] is read by one read system call. Then the
file fi is closed by a close system call. Here, every file fi read by a process pi
is different from the other processes and is the same in size. At time stime, a
process pi starts. Before performing the process pi , a pair of the memory buffer
of the virtual machine vmk and the host server are flushed. Time stime when the
open system call is issued and time etimei when the close system call is issued
by a process pi are obtained in nanosecond order by the gettimeofday system
call [13]. The execution time of the process pi is defined to be etimei − stimei
[sec]. V Tk,t stands for the average execution time of the processes pi on a virtual
machine vmk of a server st . HTh means the average execution time of the process
pi which is directly performed on the server sh .
1. First, we measure a pair of the average execution time HTh and V Tk,h of a
process pi on the server sh and on the virtual machine vmk of the server sh ,
respectively [Fig. 3(1)].
2. Next, the virtual machine vmk on the home server sh migrates to the guest
server sg . Then, after the virtual machine vmk migrates from the server sh
to the server sg , the process pi starts to manipulate the file f . The execution
time V Tk,g is measured [Fig. 3(2)].
62 N. Ogura et al.
3. The virtual machine vmk migrates from the server sh to the server sg and
then migrates back to the server sh in the live manner. Here, the process pi
is performed in the migrations and end on the server sh . Here, the execution
time V Tk,hgh is measured [Fig. 3(3)].
Figure 4 shows the average execution time HTh , V Tk,h , V Tk,g , and V Tk,hgh
[sec] of the process pi for size of the file fi . HTh is one sixth to one seventh of
the others. It takes six to seven times longer to perform a storage process on a
virtual machine than a server. On the other hands, the difference between V Tk,h ,
V Tk,g , and V Tk,hgh are small, several percentage, independently of which server
holds the virtual machine vmk .
and creates exist on the storage server s0 . The file fi can be manipulated only
through the virtual machine vmk .
A virtual machine vmk is dormant if no process is performed. Every dormant
virtual machine vmk is on the storage server s0 . A virtual machine where at least
one process is performed is active.
Suppose a client issues a process pi to a cluster S. The process pi is performed
as follows.
1. If the process pi uses an existing file fi , a home virtual machine vmk of the
file fi is found in the cluster S. Then, the process pi is issued to the virtual
machine vmk , which may be on a working server st .
2. If the process pi uses a file which does not exist in the cluster, one dormant
virtual machine vmk is taken on the storage server s0 . First, one working
server st is selected. Then, the virtual machine vmk migrates to the selected
server st . The process pi is performed on the virtual machine vmk of the host
server st .
3. An active virtual machine vmk on a working server st migrates to another
working server su if more number of processes are performed on the server st
than some value.
4. Suppose a virtual machine vmk holds a read-only file fi , i.e. vmk is static. If
more number of processes are performed on the virtual machine vmk , a clone
of the virtual machine vmk is created on the storage server s0 .
5. If a virtual machine vmk gets idle, i.e. every process terminates, the virtual
machine vmk migrates to the storage server s0 and gets dormant.
5 Concluding Remarks
In this paper, we discussed storage processes which create, read, and write files
and are performed on virtual machines. Storage processes read and write data in
files which are stored in the storage server. We proposed the model for performing
storage processes on virtual machines and migrating virtual machines. Here,
every virtual machine is initially on a storage server where every file used by the
virtual machine is stored. If a process which uses a file fi is issued to a cluster, a
virtual machine vmk holding the file fi is found. If an idle virtual machine vmk
is on a storage server, one working server is selected and one virtual machine
vmk migrates to the working server. The process is performed on the virtual
machine. If a virtual machine vmk supports only read-only files, a clone of the
virtual machine vmk is created to perform new storage processes. An active
virtual machine migrates to another working server. If every process terminates
on a virtual machine, the virtual machine backs to the storage server.
We are designing an algorithm to select a working server to which a virtual
machine migrates so that the electric energy consumption of servers [9–12] can
be reduced in a cluster.
References
1. The centos linux distribution (centos linux). https://www.centos.org/
2. Cloning virtual machines for kernel based virtual machine. https://access.
redhat.com/documentation/ja-jp/red hat enterprise linux/7/html/virtualization
deployment and administration guide/cloning virtual machines
3. A virtualization infrastructure for the linux kernel (kernel-based virtual machine).
https://en.wikipedia.org/wiki/Kernel-based Virtual Machine
4. Duolikun, D., Nakamura, S., Enokido, T., Takizawa, M.: An energy-efficient
dynamic live migration of multiple virtual machines. In: Proceedings of the 21st
International Conference on Network-Based Information Systems (NBiS-2018), pp.
87–98 (2018)
5. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: Asynchronous migration
of process replica in a cluster. In: Proceedings of the 11th International Confer-
ence on Broadband and Wireless Computing, Communication and Applications
(BWCCA-2017), pp. 161–172 (2017)
66 N. Ogura et al.
6. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: An Eco algorithm for
dynamic migration of virtual machines in a server cluster. In: Proceedings of
the 20th International Conference on Network-Based Information Systems (NBiS-
2017), pp. 42–54 (2017)
7. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: Energy-aware dynamic
migration of virtual machines in a server cluster. In: Proceedings of the 11th
International Conference on Complex, Intelligent and Software Intensive Systems
(CISIS-2017), pp. 70–81 (2017)
8. Duolikun, D., Watanabe, R., Enokido, T., Takizawa, M.: An Eco migration algo-
rithm of virtual machines in a server cluster. In: Proceedings of IEEE the 32nd
International Conference on Advanced Information Networking and Applications
(AINA-2018), pp. 189–196 (2018)
9. Enokido, T., Aikebaier, A., Takizawa, M.: A model for reducing power consumption
in peer-to-peer systems. IEEE Syst. J. 4(2), 221–229 (2010)
10. Enokido, T., Aikebaier, A., Takizawa, M.: An extended simple power consump-
tion model for selecting a server to perform computation type processes in digital
ecosystems. IEEE Trans. Ind. Inf. 10(2), 1627–1636 (2014)
11. Enokido, T., Takizawa, M.: Power consumption and computation models of virtual
machines to perform computation type application processes. In: Proceedings of
the 9th International Conference on Complex, Intelligent, and Software Intensive
Systems (CISIS-2015), pp. 126–133 (2015)
12. Kataoka, H., Duolikun, D., Enokido, T., Takizawa, M.: Multi-level computation
and power consumption models. In: Proceedings of the 18th International Confer-
ence on Network-Based Information Systems (NBiS -2015), pp. 40–47 (2015)
13. Misawa, A., Tatani, F., Kawaguchi, T.: Linux Command Book for Beginners. SB
Creative (2015)
14. Rafaels, R.J.: Cloud Computing: From Beginning to End. Create Space Indepen-
dent Publishing Platform (2015)
Energy Efficient Scheduling
of Smart Home
1 Introduction
Over the last few decades, a substantial growth has been reported in the demand
of electricity. To cope with the increasing electricity demand problem, the power
generation companies focused on the installation of new energy generators. How-
ever, installing new generation sources is neither an economical nor environment
friendly option. It is due to the fact that these sources mostly operate by fos-
sil fuels like natural gas, oil and coal. It is believed that these sources emit
carbon dioxide into the atmosphere which causes global warming. An alter-
nate solution to the intensifying electricity demand is the efficient utilization
of existing resources. To efficaciously utilize the existing resources Traditional
Grid (TG) is experiencing numerous challenges such as centralized generation
system, high power losses, aged infrastructure and wired technology etc. One of
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 67–79, 2019.
https://doi.org/10.1007/978-3-030-15035-8_7
68 S. Khan et al.
the most important challenge is the reliability and sustainability of power grid.
The increasing energy demand in industrial, residential and commercial sectors
has a high influence on the grid reliability, as a result power failures or blackout
occurs frequently [1]. The major factors that affects the reliability of TG are the
lack of communication and inefficacious load management capability. Therefore,
to make the power system more reliable and sustainable, the infrastructure of
Smart Grid (SG) is introduced.
SG is an electric grid which is revolutionized by amalgamating Information
and Communication Technology (ICT) based solutions [2]. These ICT based
solutions enable SG to have a bidirectional communication between the con-
sumers and the utility. This is achieved by amalgamating the Advanced Meter-
ing Infrastructure (AMI) and smart appliances in the existing power system.
AMI enables the consumers to learn their real time energy usage and its respec-
tive cost. The advantages of SG over TG are: reliable grid operations, dis-
tributed energy production, Demand Response (DR) strategies, Smart Meter
(SM), Demand Side Management (DSM) and Smart Scheduler (SS) [3]. DSM
plays a key role in enhancing the efficiency of SG to improve the reliability of
power system. DSM refers to the strategies adopted by the utility companies
to motivate consumers to actively participate in efficacious energy utilization
[4]. For this purpose, incentive is provided to the consumers so that energy is
efficaciously utilized. In return, the consumers have to minimize or shift the
operational time of the electric appliances to other time slots (Off-peak hours).
DR aims to persuade the consumers to modify their energy utilization pattern
in distinct hours (On-peak hours). For this purpose two types of DR policies
are introduced. These policies are incentive based DR policy and price based
DR policy. In the incentive based DR policy, consumers agree to minimize their
energy utilization in order to get some incentives from the utility. Incentive to
the consumers is given in the form of minimized electricity bill. In price based
DR scheme numerous pricing schemes are used across the world to encourage
the consumers to alter their energy utilization pattern. The well known pricing
schemes are Critical Peak Pricing (CPP), Inclined Block Rate Pricing (IBRP),
day ahead Time of Use Pricing (ToUP) and Real Time Pricing (RTP) [5].
Some of the common objectives of SG are cost minimization, optimal load
management, the reliability and sustainability of power system and Peak to
Average Ratio (PAR) reduction. In order to achieve the conflicting objectives
researchers have implemented many meta-heuristic and Linear Programming
(LP) techniques [6]. In this paper, we present an Improved Algorithm for PAR
Reduction (IAPR) in the environment of a Smart Home (SH). The motivation for
PAR minimization in residential sector is due to two reasons: (i) industrial and
commercial users are reluctant to alter their energy utilization pattern: (ii) about
40% of the total energy is consumed by the residential sector. Furthermore, about
65% of electricity consumption can be minimized by SHs and smart buildings
[7]. The main contribution of this paper are following:
Energy Efficient Scheduling of Smart Home 69
The rest of this paper is organized as follows. Literature review of the existing
schemes is described in Sect. 2. Sections 3 and 4 discusses the problem statement
and proposed system model. Scheduling techniques are presented in Sect. 5. Dis-
cussion and results are presented in Sect. 6 and finally Sect. 7 concludes the
paper.
2 Related Work
For the last few years, many DSM strategies have been developed to optimize
the energy utilization in residential sector with varying objectives.
In [8], a SHEMS is developed to monitor and schedule household electrical
appliances in a conventional house to minimize electricity consumption cost. In
this scheme, if the load on the grid exceeds a predefined limit, the new appli-
ances are either shifted to other time periods or operated using battery power.
Batteries are either charged from the utility in the low price intervals or using
Photovoltaic (PV) cells in day time. Experimental results demonstrate that the
scheme reasonably minimizes electricity consumption cost. However, this work
did not consider the various pricing schemes. Yang et al. [9] present a scheme
using Improved Particle Swarm Optimization (IPSO) for scheduling home appli-
ances. The scheme mainly focuses on minimizing cost. It is evident that elec-
tricity tariff and objective curve have an inverse relationship with each other.
Experimental results show that the desired objective curve is achieved. More-
over, this work aims to achieve power system stability. For this purpose, the
high electricity consumption is reduced in high price intervals. As a result, UC
is compromised.
The authors in [10] propose a Mixed Integer Linear Programming (MILP)
based cost minimization and load equalizing scheme for domestic users. In this
scheme, electricity consumer is considered as a prosumer. Prosumer is the pro-
ducer as well as the consumer of electricity. The work considers that prosumers
have installed PV cells in their homes for electricity generations along with a
connection to a commercial grid in order to fulfill their energy requirements.
In the proposed work, prosumers store energy in batteries only when electricity
generation exceeds the user demand. However, this work ignores exporting sur-
plus energy to maximize their income. The authors in [11] develop a fuzzy based
energy management system. In this work, consumer is involved in changing the
status of appliances from ON-state to OFF-state. The work mainly focuses on
minimizing cost along with PAR. The UC in this scheme is compromised due to
increase in the appliances waiting time.
Israr et al. [12] present an optimization function to maximize the UC. The
proposed scheme aims to retain an intended environment inside a smart home.
Genetic Algorithm (GA) and PSO techniques are used to enhance the UC. For
70 S. Khan et al.
this purpose, the recorded data for a month in an indoor laboratory environment
is analyzed. Experimental results demonstrate that the proposed scheme achieves
desired results. However, this work did not consider electricity consumption cost
and PAR. In [13], a hybrid technique using Harmony Search Algorithm (HSA)
and Enhanced Differential Evaluation (EDE) algorithm is developed. The pro-
posed scheme formulate the demand and supply problem as Multiple Knapsack
Problem (MKP). The work incorporates ESS to maximize the UC and alleviate
PAR. Experimental results demonstrate that the hybrid scheme reasonably min-
imizes the consumer bill as compare to HSA and EDE. However, PAR reduction
using EDE is better as compared to their proposed technique. A chance con-
strained DR scheme is proposed in [14] to schedule electrical appliances in HEMS.
The scheme uses PSO and two-Point Estimation Method (2-PEM). Experimen-
tal results demonstrate that PSO along-with 2-PEM outperforms gradient based
PSO with Latin hyper cube sampling in terms of minimizing computational load.
However, this work, ignores PAR and consumption cost. PAR burdens the utility
that has a direct impact on consumer.
Wen et al. [15] propose HEMS using reinforcement learning technique to
meet the DR of residential buildings. The scheme adds intelligence to the HEMS
via a three step Observe, Learn and Adopt algorithm (OLA). In this technique,
an algorithm learns the electricity consumption pattern of the appliances based
on consumer preferences. Experimental results demonstrate that the proposed
scheme reasonably minimizes the user consumption cost and PAR. However, UC
is compromised due to shifting or delaying the operational time of the electricity
consuming appliances. The work in [16] minimizes PAR in SG via DR strategies.
This work considers the optimization problem as two stage problem. The con-
sumers try to boost their pay-offs and the utility tries to maximize its revenue.
At the consumer end, electricity consumption cost is calculated using an itera-
tive algorithm. Whereas, at the retailer end, the work used Simulated-Annealing
based Price Control Algorithm (SAPCA). SAPCA is used to maximize retailer’s
profit using RTP scheme. Experimental results demonstrate that the proposed
scheme is beneficial for the consumers as well as utility. Furthermore, the pro-
posed work reduces the PAR.
Farrokhifar et al. in [17] develop a scheme for scheduling home appliances to
reduce consumption cost within a building. The scheme considered optimization
problem as the ILP problem. Simulations results demonstrate that the proposed
scheme reduces consumption cost. However, this work ignores the PAR. Khan
et al. in [18] present a priority induced DSM scheme to efficaciously schedule
the home appliances in SHEMS. The main objective of this work is to mini-
mize cost and mitigate rebound peaks. In this work, the authors used renowned
meta-heuristic techniques namely EDE, BPSO, and GA. Furthermore, knapsack
capacity limit is used to enhance the grid reliability. Simulation results shows
that this scheme minimized the consumer bill by 60%. However, this work con-
sidered a limited number of home appliances.
Energy Efficient Scheduling of Smart Home 71
4 System Model
This section discusses the proposed system model. In this paper, the aim is to
shift the execution time of home appliances such that PAR and cost is mini-
mized. A detailed description of our proposed system model is shown in Fig. 1.
To meet the energy demand, it is assumed that the SH is connected to util-
ity using AMI. AMI allows bi-directional communication between the consumer
and utility. Generally AMI comprises of a SM, energy management controller and
data management system. The SM acts as a communication gateway between
the consumer and the utility.
Utility
Smart meter
Here CS and CU represents the energy utilization cost in the scheduled and
unscheduled scenario. In order to compute the PAR, Eq. 2 is used.
LP eak
P AR = (2)
LAverage
Here LP eak is the maximum load consumed by the home appliances in SH
for a single time slot and LAverage is the average load consumed in one day. To
compute the per hour and total load consumed by the SH we use Eqs. 3 and 4.
24
LoadT = LoadH (t) (4)
t=1
Here LoadH represents the load consumed in one hour, LoadT is the total
load consumed in one day, ρ(t) shows the power consumed by an appliance and
α is the ON and OFF status of an appliance.
5 Scheduling Techniques
This section discusses the existing and proposed scheduling schemes.
Energy Efficient Scheduling of Smart Home 73
5.3 IAPR
It is evident that plant survives by their vegetative propagation nature. For
their purpose, plants send runners and root hair randomly. If the runners or the
74 S. Khan et al.
4
PAR
0
Unscheduled SA SSA IAPR
4
PAR
0
Unscheduled SA SSA IAPR
Figures 2 and 3 represents that IAPR outperforms the well known meta-
heuristic techniques SSA and SA. The proposed IAPR minimized PAR using
CPP by 72%. Whereas, SA and SSA minimized PAR by 51% and 9% respectively.
In RTP, PAR alleviation is 17%, 48% and 68% in SSA, SA and IAPR respectively.
Figures 4 and 5 shows the per hour load consumption pattern of SH. These
figures gives a complete picture of the performance of IAPR using CPP and
RTP. The IAPR balanced the per hour load consumption, as a result PAR in
minimized. The maximum load recorded in a single time slot using CPP and RTP
schemes is 10.54, 7.80, 9.60 and 7.80 kWh for unscheduled, SA, SSA and IAPR
respectively. Here it is worth mentioning that PAR is minimized by balancing
load consumption throughout the day and total load consumed by all algorithms
in both of the pricing scheme was same.
76 S. Khan et al.
12
Unscheduled SA SSA IAPR
4
2
0
1 4 8 12 16 20 24
Time (hours)
12
Per hour load (kWh)
8
6
2
0
1 4 8 12 16 20 24
Time (hours)
1200
Unscheduled
Per hour cost (cents)
1000 SA
SSA
800 IAPR
600
400
200
0
1 4 8 12 16 20 24
Time (hours)
200 Unscheduled
100
50
0
1 4 8 12 16 20 24
Time (hours)
5000
Totalcost(cents)
4000
3000
2000
1000
0
Unscheduled SA SSA IAPR
1500
Total cost (cents)
1000
500
0
Unscheduled SA SSA IAPR
The waiting time of all the algorithms using CPP and RTP scheme is shown
in Figs. 10 and 11. The proposed IAPR has a higher waiting time as compare to
the other optimization algorithms. It can be seen from Figs. 2, 3, 4, 5, 6, 7, 8, 9,
10 and 11 that all the algorithms with both the pricing strategies are confronted
with trade-offs. SSA minimized the electricity consumption cost. However, the
performance in PAR minimization is not good and multiple peaks are created
in Off-peak hours. SA minimized the PAR as compare to SSA. However its
performance in waiting time reduction is higher than SSA. Similarity IAPR
minimized PAR, however, it increased the appliances waiting time as compare
to SA and SSA.
78 S. Khan et al.
0
Unscheduled SA SSA IAPR
7
6
Waiting time (hours)
5
4
3
2
1
0
Unscheduled SA SSA IAPR
References
1. Nadeem, J., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
I.A.: A hybrid genetic wind driven heuristic optimization algorithm for demand
side management in smart grid. Energies 10(3), 319 (2017)
2. Farhangi, H.: The path of the smart grid. IEEE Power Energy Mag. 8(1) (2010)
3. Hussain, B., Javaid, N., Hasan, Q., Javaid, S., Khan, A., Malik, S.: An inventive
method for eco-efficient operation of home energy management systems. Energies
11(11), 3091 (2018)
Energy Efficient Scheduling of Smart Home 79
4. Khalid, A., Javaid, N., Mateen, A., Ilahi, M., Saba, T., Rehman, A.: Enhanced
time-of-use electricity price rate using game theory. Electronics 8(1), 48 (2019)
5. Economics, Frontier, and Sustainability First: Demand side response in the domes-
tic sector-a literature review of major trials. Final report, London, August 2012
6. Shareef, H., Ahmed, M.S., Mohamed, A., Al Hassan, E.: Review on home energy
management system considering demand responses, smart technologies, and intel-
ligent controllers. IEEE Access 6, 24498–24509 (2018)
7. Lior, N.: Sustainable energy development: the present (2009) situation and possible
paths to the future. Energy 35(10), 3976–3994 (2010)
8. Shakeri, M., Shayestegan, M., Reza, S.S., Yahya, I., Bais, B., Akhtaruzzaman, M.,
Sopian, K., Amin, N.: Implementation of a novel home energy management sys-
tem (HEMS) architecture with solar photovoltaic system as supplementary source.
Renewable Energy 125, 108–120 (2018)
9. Yang, H.-T., Yang, C.-T., Tsai, C.-C., Chen, G.-J., Chen, S.-Y.: Improved PSO
based home energy management systems integrated with demand response in a
smart grid. In: 2015 IEEE Congress on Evolutionary Computation (CEC), pp.
275–282. IEEE (2015)
10. Van Der Stelt, S., AlSkaif, T., van Sark, W.: Techno-economic analysis of household
and community energy storage for residential prosumers with smart appliances.
Appl. Energy 209, 266–276 (2018)
11. Collotta, M., Pau, G.: A novel energy management approach for smart homes using
bluetooth low energy. IEEE J. Sel. Areas Commun. 33(12), 2988–2996 (2015)
12. Ullah, I., Kim, D.H.: An improved optimization function for maximizing user com-
fort with minimum energy consumption in smart homes. Energies 10(11), 1818
(2017)
13. Kazmi, S., Javaid, N., Mughal, M.J., Akbar, M., Ahmed, S.H., Alrajeh, N.: Towards
optimization of metaheuristic algorithms for IoT enabled smart homes targeting
balanced demand and supply of energy. IEEE Access (2017)
14. Huang, Y., Wang, L., Guo, W., Kang, Q., Qidi, W.: Chance constrained optimiza-
tion in a home energy management system. IEEE Trans. Smart Grid 9(1), 252–260
(2018)
15. Wen, Z., O’Neill, D., Maei, H.: Optimal demand response using device-based rein-
forcement learning. IEEE Trans. Smart Grid 6(5), 2312–2324 (2015)
16. Qian, L.P., Zhang, Y.J.A., Huang, J., Wu, Y.: Demand response management via
real-time electricity price control in smart grids. IEEE J. Sel. Areas Commun.
31(7), 1268–1280 (2013)
17. Farrokhifar, M., Momayyezi, F., Sadoogi, N., Safari, A.: Real-time based approach
for intelligent building energy management using dynamic price policies. Sustain.
Cities Soc. 37, 85–92 (2018)
18. Khan, A., Javaid, N., Ahmad, A., Akbar, M., Khan, Z.A., Ilahi, M.: A priority-
induced demand side management system to mitigate rebound peaks using multiple
knapsack. J. Ambient Intell. Humanized Comput., 1–24 (2018)
19. Pooranian, Z., Abawajy, J.H., Conti, M.: Scheduling distributed energy resource
operation and daily power consumption for a smart building to optimize economic
and environmental parameters. Energies 11(6), 1348 (2018)
20. Merrikh-Bayat, F.: A numerical optimization algorithm inspired by the strawberry
plant. arXiv preprint arXiv:1407.7399 (2014)
21. Mirjalili, S., Gandomi, A.H., Mirjalili, S.Z., Saremi, S., Faris, H., Mirjalili, S.M.:
Salp Swarm Algorithm: a bio-inspired optimizer for engineering design problems.
Adv. Eng. Softw. 114, 163–191 (2017)
Minimizing Daily Cost and Maximizing
User Comfort Using a New Metaheuristic
Technique
1 Introduction
In recent decades an exponential growth in the energy consumption is observed,
therefore, the gap between demand and generation is increasing. Because of the
rise in the energy need, the conventional grid is facing a number of problems
like reliability, maintenance and sustainability. Residential consumer are grow-
ing rapidly at a pace of 0.3% annually, and has large share of 27% among others
in electricity consumption [1]. In past most of the power generation was achieved
using fossil fuels to fulfill the gap between energy demand and generation. How-
ever in recent years scientists have worked on exploring new means of energy
generation, i.e., renewable and sustainable energy resources [2]. The addition of
Renewable Energy Sources (RESs) resulted in an increase in the power system
complexity, where existing grids were not able to manage that. Smart Grid (SG)
is introduced to handle the above mentioned challenges. A SG is equipped with
Information and Communication Technology (ICT), that enabled it to incorpo-
rate RESs and manage the stability and reliability of power systems. The most
important feature of SG is the control of energy production, transmission and
distribution through advanced ICTs.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 80–92, 2019.
https://doi.org/10.1007/978-3-030-15035-8_8
HMS Using Hybrid Meta-heuristic Optimization Technique 81
The state of the art techniques for DSM and SG are discussed in Sect. 2.
Different problems related to the DR and DSM are reflected along with their
consequences in Sect. 3. Section 4 explains the suggested scheme for discussed
problemettes. Section 5 highlights formerly employed techniques, Strawberry
Algorithm (SBA), Earthworm optimization Algorithm (EWA), and schemed
algorithm. Simulations outcomes are depicted and shown in Sect. 6. At last,
Sect. 7 draws conclusion of the work done.
2 Related Work
3 Problem Statement
From the time of its evolution, SG has a number of complications as well, i.e.,
energy transferring, equality, surveillance and secrecy at DSM level [5]. The
uneven use of electricity on user side rises the burden on the power distribution
and power generating system. Due to irregular usage, energy required in some
hours (peak hours) increases exponentially. Such high demand results in using
extra generation units to provide the required energy, which subsequently results
increase in cost. Therefore, DSM needs a proficient scheme that can distribute
the load between peak hours and off peak hours evenly.
Although, authors have already done a lot of work on DSM for reducing cost,
waiting time and PAR. Authors in [6,10,17,18] worked on DR optimization con-
sidering uncertainties of must run load, aimed at reducing the upper bound of
energy consumption cost, however PAR was not considered. In [7,8,11] authors
worked on reducing the peak demand through scheduling techniques for appli-
ances incorporating the RES, however user comfort was not considered.
energy to SM. SM sends the information coming through the utility with the
EMC. Now, EMC possesses the data collected from appliances and the cost infor-
mation received from utility. Information collected from utility and appliances
in used by EMC for optimum appliances scheduling. EMC uses the Strawberry
Algorithm (SBA), Earthworm Optimization Algorithm (EWA) and proposed
scheme for scheduling. The goal of the scheduling is to reduce PAR, maximize
user satisfaction through minimizing waiting time and reducing cost. The dis-
cussed system model is displayed in Fig. 1.
In our proposed model, we have a house, using 11 appliances and those appli-
ances are distributed among different classes. We classify appliances in three
main classes, i.e., Fixed, shift-able and interruptible appliances. Fixed appli-
ances require to be ON according to the specified schedule. Those appliances
that are optional to scheduling but the can not be interrupted while they are
working are classified as Shiftable appliances. Furthermore, there is a condition
for washing dryer that it can start its operation when washing machine has com-
pleted its operation. Appliances those are schedulable as well as interruptible are
placed in Interruptible appliances category. Table 1, displays the classified smart
appliances as well as their LoT and power rating in respective columns. The pro-
posed schemes objective is to: minize PAR, maximize user comfort by reducing
waiting time and reducing total cost.
5.1 SBA
SBA is a nature inspired numerical optimization algorithm. This algorithm is
derived from the strawberry plant behavior for searching the resources and is
being used for solving complex computing problems. Strawberry plant uses both
runners and root for searching resources like water and minerals. Where, these
runners being used for global search and roots being used for local search. Runner
and roots are randomly initialized at the beginning. While algorithm running,
when these runners and roots finds new resources then the reproduction starts
which results in more runners and roots. The locations of roots and runners at
a specific iteration is calculated using Eq. 4. Where, Xroot (i) and Xrunner (i) are
matrices containing locations of roots and runners respectively. Equations 5 and
6 are used for optimization and is the mathematical representation of strawberry
algorithm.
5.2 EWA
Earthworm executes numerous optimization runs during the reproduction pro-
cess. Reproduction process of the earthworm follows following steps.
• Every earthworm is capable of participating in regeneration process, it can
reproduce at most two earthworms.
• The singular generated earthworm contains the complete genetic behavior of
its parent that is of same length as parent.
• The single fit earthworm supports a straight next generation and can not be
changed by operators. It guarantees that earthworm can not end with the
increment in generations.
HMS Using Hybrid Meta-heuristic Optimization Technique 85
5.3 Proposed
Our proposed scheme is based on the behavior of plant roots and earthworm.
Plant root search resources, i.e., minerals and water and grows in the direction of
best possible route to the resource. Earthworm performs numerous runs during
the reproduction process. Every earthworm has the ability to reproduce and has
same capability of reproduction like parent earthworm. In our proposed scheme
we consider plants as solution to the problem, plant roots for finding the global
optimum solution and earthworm capability for local optimum solution. The
resource found quality is mapped as fitness function in our proposed scheme.
The steps followed in our proposed scheme are listed below.
Washing machine, cloth dryer and dish washer are included in SA. These appli-
ances can not be interrupted once they have started their working, however these
appliances can be shifted. Water heater, space heater, iron and vacuum cleaner
are included in IA. This group of appliances can be shifted as well as interrupted.
CPP and RTP are used for load, cost and PAR calculation.
The PAR for unscheduled, SBA, EWA and proposed scheme using CPP
and RTP pricing schemes is displayed in Figs. 8 and 9 respectively. In case of
CPP pricing scheme PAR is reduced by 58.77%, 62.29% and 58.15% using SBA,
EWA and proposed scheme. Here, EWA outperformed other schemes. In case of
RTP pricing scheme PAR is reduced by 54.99%, 54.86% and 51.42% using SBA,
EWA and proposed scheme. Here, SBA outperformed other schemes. It is worth
mentioning that all schemes performed better than unscheduled.
Figures 10 and 11 are displaying the waiting time for unscheduled, SBA, EWA
and proposed scheme using CPP and RTP pricing schemes respectively. In case
of CPP pricing scheme waiting time is reduced to 23.37%, 25.41% and 18.77% for
SBA, EWA and proposed scheme. Here, proposed scheme outperformed other
schemes. In case of RTP pricing scheme waiting time is reduced to 24.16%,
24.28% and 18.13% using SBA, EWA and proposed scheme. Here, proposed
scheme outperformed other schemes. It is also noticed that all schemes performed
better than unscheduled.
88 R. A. Abbasi et al.
Figures 12 and 13 are displaying the total cost for unscheduled, SBA, EWA
and proposed scheme using CPP and RTP pricing schemes respectively. In case
of CPP pricing scheme cost is reduced by 18.75%, 30.65% and 26.99% using
SBA, EWA and proposed scheme. Here, EWA outperformed other schemes. In
case of RTP pricing scheme cost is reduced by 13.99%, 15.25% and 27.05% using
SBA, EWA and proposed scheme. Here, our proposed scheme outperformed other
schemes.
The proposed scheme minimizes the waiting time, however, it compromises
for PAR and cost in case of CPP pricing signal. Where, proposed scheme mini-
mizes the waiting time and cost, however, it compromises PAR in case of RTP
pricing signal.
90 R. A. Abbasi et al.
7 Conclusion
DSM is focused to manage the appliances on the user end for efficient use of
energy. Load shifting technique of DSM is used in proposed HEMS to attain
the desired results. A single house consisting of various appliances is considered
for scheduling. Each appliance of house is scheduled using three heuristic tech-
niques; SBA, EWA, and proposed scheme. The proposed scheme helps to find the
most optimal schedule of each house appliance considering the constraints. The
performance of applied techniques is assessed with respect to basic parameters:
PAR, user comfort and energy consumption cost. The cost of energy consump-
tion is calculated using RTP and CPP pricing schemes. Control parameters and
categorization of appliances is kept same, however, power rating of electrical
devices and their types is kept same for a fair comparison. Results are depicting
that waiting time is reduced to 18.77% in case of CPP and 18.13% for RTP
pricing scheme which is best with respect to SBA and EWA. Cost is reduced by
27.05% in case of RTP which is best in comparison to the other two techniques.
However, there is a trade-off, as we cannot achieve the best values for all three
performance parameters simultaneously.
References
1. U.S. Energy Information Administration, Annual Energy Outlook 2018, p. 121,
122, Residential and commercial energy consumption grows gradually from 2017
to 2050. https://www.eia.gov/outlooks/aeo/pdf/AEO2018.pdf. Accessed October
2018
2. Luo, F., Ranzi, G., Wan, C., Xu, Z., Dong, Z.Y.: A multi-stage home energy
management system with residential photovoltaic penetration. IEEE Trans. Ind.
Inf., 11 (2018)
3. Strbac, G.: Demand side management: benefits and challenges. Energy Policy 36,
4419–4426 (2008). Energies 11, 1426 (2018)
4. Gellings, C.W.: The concept of demand-side management for electric utilities. Proc.
IEEE 73(10), 1468–1470 (1985)
5. Baharlouei, Z., Hashemi, M.: Demand side management challenges in smart grid:
a review. In: Proceedings Smart Grid Conference (SGC), pp. 96–101, December
2013
6. Huang, Y., Wang, L., Guo, W., Kang, Q., Wu, Q.: Chance constrained optimization
in a home energy management system. IEEE Trans. Smart Grid 9(1), 252–260
(2018)
7. Hansen, T.M., Chong, E.K.P., Suryanarayanan, S., Maciejewski, A.A., Siegel, H.J.:
A partially observable markov decision process approach to residential home energy
management. IEEE Trans. Smart Grid 9(2), 1271–1281 (2018)
8. Luo, F., Ranzi, G., Kong, W., Dong, Z.Y., Wang, F.: Coordinated residential energy
resource scheduling with vehicle-to-home and high photovoltaic penetrations. IET
Renew. Power Gener. 12(6), 625–632 (2018)
9. Luo, F., Dong, Z.Y., Xu, Z., Kong, W., Wang, F.: Distributed residential energy
resource scheduling with renewable uncertainties. IET Gener. Transm. Distrib.
12(11), 2770–2777 (2018)
92 R. A. Abbasi et al.
10. Elghitani, F., Zhuang, W.: Aggregating a large number of residential appliances for
demand response applications. IEEE Trans. Smart Grid 9(5), 5092–5100 (2018)
11. Tushar, M.H.K., Zeineddine, A.W., Assi, C.: Demand-side management by regu-
lating charging and discharging of the EV, ESS, and utilizing renewable energy.
IEEE Trans. Ind. Inf. 14(1), 117–126 (2018)
12. Wu, X., Hu, X., Yin, X., Moura, S.J.: Stochastic optimal energy management of
smart home with PEV energy storage. IEEE Trans. Smart Grid 9(3), 2065–2075
(2018)
13. Moon, J., Kim, K.-H., Kim, Y., Hwang, E.: A short-term electric load forecasting
scheme using 2-stage predictive analytics. In: 2018 IEEE International Conference
on Big Data and Smart Computing (2018)
14. Keles, D., Scelle, J., Paraschiv, F.: Wolf Fichtner Extended forecast methods for
day-ahead electricity spot prices applying artificial neural networks. Appl. Energy
162, 218–230 (2016)
15. Ghofrani, M., Hassanzadeh, M., Etezadi-Amoli, M., Fadali, M.: Smart meter based
short-term load forecasting for residential customers. In: North American Power
Symposium (NAPS), pp. 1–5, August 2011
16. Huang, H., Cai, Y., Xu, H., Yu, H.: A multi-agent minority-game- based demand-
response management of smart buildings towards peak load reduction. IEEE Trans.
Comput.-Aided Des. Integr. Circuits Syst. 36(4), 573–585 (2017)
17. Ruelens, F., Claessens, B.J., Vandael, S., Schutter, B.D., Babuška, R., Belmans,
R.: Residential demand response of thermostatically controlled loads using batch
reinforcement learning. IEEE Trans. Smart Grid 8(5), 2149–2159 (2017)
18. Pourmousavi, S.A., Patrick, S.N., Nehrir, M.H.: Real-time demand response
through aggregate electric water heaters for load shifting and balancing wind gen-
eration. IEEE Trans. Smart Grid 5(2), 769–778 (2014)
19. Jindal, A., Dua, A., Kaur, K., Singh, M., Kumar, N., Mishra, S.: Decision tree
and SVM-based data analytics for theft detection in smart grid. IEEE Trans. Ind.
Informat. 12(3), 1005–1016 (2016)
20. Mirowski, P., Chen, S., Ho, T.K., Yu, C.-N.: Demand forecasting in smart grids.
Bell Labs Tech. J. 18(4), 135–158 (2014)
21. Wijaya, T.K., Vasirani, M., Humeau, S., Aberer, K.: Cluster-based aggregate fore-
casting for residential electricity demand using smart meter data. In: IEEE Inter-
national Conference on Big Data (Big Data)
22. Yang, J., Liu, J., Fang, Z., Liu, W.: Electricity scheduling strategy for home energy
management system with renewable energy and battery storage: a case study. IET
Renew. Power Gener. 12(6), 639–648 (2018)
23. Merrikh-Bayat, F.: A numerical optimization algorithm inspired by the strawberry
plant, arXiv preprint arXiv:1407.7399 (2014)
24. Waterloo North Hydro Inc. Residential Charges and Rates. Accessed 20 May 2016.
https://www.wnhydro.com/en/yourhome/time-of-use-rates.asp
Efficiency Analysis of TFHE Fully
Homomorphic Encryption Software Library
Based on GPU
Abstract. There are a large number of independent matrix and vector opera-
tions in the lattice-based homomorphic encryption. These operations are suitable
for GPU, which can greatly improve the efficiency of homomorphic operations.
In this paper, we analyze the structure of the homomorphic encryption algorithm
and verify the reliability of the homomorphic encryption software library, debug
and analyze the fully homomorphic encryption software library TFHE and its
corresponding GPU version cuFHE, and then compare their efficiency. The
experimental results show that the GPU version TFHE is 4.5 times faster than
the CPU version TFHE, so the GPU can greatly improve the homomorphic
running speed of the homomorphic encryption scheme.
1 Introduction
The FHE (fully homomorphic encryption) algorithm can solve the problem of data
security and privacy in cloud computing [1]. In cloud computing, the data owner
uploads the data to the server for processing, but the server does not consider the data
to be trusted, so the data needs to be encrypted before uploading. Data encrypted by
traditional data encryption will lose its original characteristics, causing most data
processing methods to fail [2]. FHE can directly operate on ciphertext, and the result of
the operation is equal to the result of doing the same operation on plaintext [3].
In 2009, Gentry proposed the first FHE scheme [4]. In this scheme, Gentry pro-
posed a construction idea: Bootstrapping (running its own decryption circuit with the
homomorphic scheme) + Squashing (squashing the depth of the decryption circuit),
which is also known as the Gentry blueprint.
In order to apply the homomorphic algorithm to the actual situation and improve the
efficiency of the scheme, the researchers continue to improve the bootstrap process [5].
At CRYPTO2009, Gentry, Sahai and Waters proposed a GSW [6] FHE scheme based
on approximate eigenvectors, which uses Cipher Flatten technology to construct a rel-
atively simple homomorphic encryption scheme based on LWE. At ASIACRYPT2016,
Chillotti uses the feature that the product is not greater than the multiplier on the multi-
plication of T = (0,1], constructs a more efficient homomorphism scheme based on the
variant TGSW [7]. It replaces the product of the original TGSW ciphertext (matrix) and
the TGSW ciphertext (matrix) by the external product of TGSW ciphertext (matrix) and
TLWE ciphertext (vector), and constructs the addition on polynomial exponents more
efficiently, which reduce the time of the bootstrapping process to 0.1 s. At ChinaCrypt
2017, Zhou [8] optimized the accumulating operation in the bootstrapping process, which
increased the speed of the single-bit bootstrapping process to 10 ms/gate, and it is the
most efficient single-bit bootstrap FHE scheme.
Recently, GPUs have been widely used in general-purpose computing. The CUDA
enables GPU parallel computing to have better programmability, and programming is
more efficient and flexible. This advantage is also used to further improve the efficiency
of FHE. In 2009, Wang et al. implemented megabit homomorphic multiplication by
GPU parallel computing of FFT and Barrett modular reduction algorithm [9, 10]. In
2015, Wei et al. implemented a GPU-accelerated NTRU scheme by modulo multi-
plication in the remaining space, and processed the large-coefficient polynomial by the
NTT and CRT methods, thereby realizing the GPU optimization of the LTV hierar-
chical homomorphic encryption scheme [11, 12].
TFHE is currently a more efficient single-bit FHE algorithm. In order to reduce the
noise, the scheme reduces the range of the plaintext to (0,1] during the bootstrap
process, and decomposes the ciphertext. In this paper, we debug the TFHE library and
its corresponding GPU accelerated version cuFHE, test its operation mode, and
dynamically compare and analyze their performance. Our goal is to verify the relevant
properties of FHE and analyze the related properties of the algorithm from the per-
spective of program implementation, providing a practical basis for GPU accelerated
homomorphic encryption research.
2 Preliminary
2.1 Algorithm
Let the security parameter to be k, set B , f0; 1g, real number ring with a modulus of 1
T , R=Z 2 ½0; 1Þ, polynomial ring R , Z½X=ðX N þ 1Þ, polynomial ring with a mod-
ulus of 1 TN ½X , R½X=ðX N þ 1Þ, where the 2N-th cyclotomic polynomial
Q
U2N ðXÞ ¼ ðX xi2N Þ ¼ X N þ 1, Mp;q ðEÞ is a matrix size of p q that E
i2Z2N
encrypted.
Bootstrapping: h 2 Mðk þ 1Þl;k þ 1 ðTN ½XÞ, Dech;b;e is the decomposition function, s 2
Bn ; s00 2 BN ½Xk ; a, c is the noise parameter, BK ¼ BKs!s00 ;a is the bootstrapping key,
Coefficient matrix is s0 ¼ KeyExtractðs00 Þ 2 BkN , exchange key is KS ¼ KSs0 !s;c;t ,
LWE ciphertext ða;bÞ 2 LWEs;g ðlÞ, where l 2 T, two related messages l0 and l1 ,
output LWEs ðl0 Þ after bootstrapping. If jus ða; bÞj 14 d, l0 ¼ l0 ; else if
jus ða; bÞj 14 þ d, l0 ¼ l1 , where d is the accumulated deviation, the worst case is n4N
þ1
,
an the best case is 0. If #BK ¼ VarðErrðBKi ÞÞ ¼ p a , VKS ¼ VarðErrðKSi ÞÞ ¼ p c , v
2 2 2 2
SM Instruction Fetch/Dispatch
Shared Memory
TEX SP SP
SP SP
SFU SFU
SM SP SP
SP SP
The GPU can be used to speed up the FHE bootstrapping process because it
contains thousands of CUDA cores, which allows the graphics card to compute high-
dimensional matrices in parallel. In SIMT mode, the next thread starts the I/O operation
when the previous thread performs the operation. It is similar to the instruction pipe-
line. When the number of threads is enough, the large amount of overhead caused by
I/O can be blocked.
Thread management is an important part of CUDA programming. In CUDA, a
stream represents a GPU operation queue. By default, CUDA creates a default stream
that performs a series of operations in sequence. Figure 2 shows a typical GPU thread
view.
Grid
Step 6: Add the search directory to the CMAKELIST.TXT file. Insert into
CMAKELIST.TXT file: set(CMAKE_INCLUDE_PATH “/usr/local/fftw/include/”), set
(CMAKE_LIBRARY_PATH “/usr/local/fftw/lib/”) and include_directories(“/usr/local/
fftw/include/”)
Step 7: Add execution links and commands to the CMAKELIST.TXT file. add_
executable(test-bootstrapping-${FFT_PROCESSOR}testbootstrapping.cpp${TFHE_-
_HEADERS} and target_link_libraries(test-bootstrapping-${FFT_PROCESSOR}
${RUNTIME_LIBS}).
Begin
Command: make
Compile
Plaintext: in_message
Bootstrapping...
Begin
Command: ./ test_api_gpu
GPU initialization
5 Performance Analysis
In order to analyze the performance of the program, we constantly adjust the parameters,
compare the experimental results and analyze the reasons. The ciphertext of the fully
homomorphic scheme will increase the noise after homomorphic operation, so we use
bootstrap technology to reduce the noise of ciphertext to a reasonable range. In the
bootstrap process of TFHE, in order to achieve homomorphic decryption, it is necessary
to complete the homomorphic rounding function round, exponential homomorphic
multiplication and homomorphic addition. In the homomorphic constant multiplication,
the TGSW ciphertext is skillfully constructed, and the most time-consuming is just the
Efficiency Analysis of TFHE Fully Homomorphic Encryption Software Library 99
In order to verify the correctness of the scheme, we analyzed the error function erf
values of different decomposition parameters. It can be seen from Fig. 6 that the
smaller the component value 1, the larger the error function erf value and the higher the
correctness. For the same 1, when the number of decomposition bits bk < 11, the
correct rate does not change much, but when bk > 11, the correct rate will drop sharply.
Therefore, in order to ensure the correctness, the selected component value 1 cannot be
greater than 11.
Under the same hardware conditions, we test TFHE and cuFHE separately. The test
sample is 256 groups. The test time is the average time within the group. The time-
consuming of the key generation algorithm, encryption algorithm and decryption
algorithm, and the average Nand operation time are shown in Table 1. It can be seen
that the encryption and decryption algorithm of cuFHE has been significantly
improved, and the Nand gate operation time on the GPU platform is approximately 4.5
times that on the CPU platform.
We also tested the time consumption of the TFHE and cuFHE decryption algo-
rithms. The decryption time of both schemes is ms level, and the difference can be
ignored. The time consumption of uFHE’s key generation algorithm is much higher
than that of TFHE. The main reason is that the GPU is only suitable for tasks with less
branches, large amounts of data, and delay-insensitive tasks. However, the key gen-
eration algorithm involves more logical branches, so the computational efficiency is
affected. On the other hand, GPU access memory has a large latency and requires high
data alignment. A single Warp accesses a group of memory, which can only read
consecutive and aligned 128 bits at a time. But if the memory accessed is discrete, the
Efficiency Analysis of TFHE Fully Homomorphic Encryption Software Library 101
efficiency is greatly reduced. Because a single Warp core cannot perform subsequent
operations when the data is not processed. A large number of memory read and write
operations are involved in the key generation algorithm, and the process of generating
the public key and private key depends on the synchronization of the results, which
resul in a slower speed.
In Table 1, we can see that the Nand gate operation time on the GPU platform is
approximately 4.5 times that on the CPU platform. Stream management and memory
management are important factors affecting the efficiency of the algorithm. In cuFHE,
additional streams are created, allowing kernel and memcpy operations to run in par-
allel, so the execution order is more flexible and improves execution efficiency. At the
same time, use the following methods to optimize data transfer efficiency: reduce the
amount and frequency of data transmitted between the host and the device; use page-
locked memory for higher bandwidth; reasonably arrange memory space, such as
frequently used data put into the cache, reduce data cross-thread exchange.
6 Conclusion
In this paper, we debugged the homomorphic encryption software library TFHE and its
corresponding GPU version cuFHE through experiments, verified their reliability and
analyzed their performance. The experimental results show that the speed of cuFHE is
about 4.5 times that of TFHE. The GPU can greatly improve the homomorphic running
speed of FHE algorithm.
References
1. Tang, D.-H., Zhu, S.-X., Wang, L., et al.: Fully homomorphic encryption scheme from
RLWE. J. Commun. 35(1), 173–182 (2014)
2. Huang, H., Gong, T., Chen, P., et al.: Secure two-party distance computation protocol based
on privacy homomorphism and scalar product in wireless sensor networks. Tsinghua Sci.
Technol. 21(4), 385–396 (2016)
3. Chen, Z.-G., Wang, J., Song, X.-X.: Survey on fully homomorphic encryption. Appl. Res.
Comput. 31(6), 1624–1630 (2014)
4. Gentry, C.: Fully homomorphic encryption using ideal lattices. In: STOC 2009, vol. 9,
pp. 169–178 (2009)
5. Zhou, T., Yang, X., Zhang, W., Wu, L., et al.: Efficient fully homomorphic encryption with
circularly secure key switching process. a a 1, 1
6. Gentry, C., Sahai, A., Waters, B.: Homomorphic encryption from learning with errors:
conceptually-simpler, asymptotically-faster, attribute-based. In: Advances in Cryptology–
CRYPTO 2013, pp. 75–92. Springer, Heidelberg (2013)
7. Chillotti, I., Gama, N., Georgieva, M., et al.: Faster fully homomorphic encryption:
bootstrapping in less than 0.1 s. In: Advances in Cryptology–ASIACRYPT 2016:
Proceedings of the 22nd International Conference on the Theory and Application of
Cryptology and Information Security, Part I, vol. 10031, Hanoi, Vietnam, 4–8 December
2016, pp. 3–33. Springer, Heidelberg (2016)
102 H. Yang et al.
8. Zhou, T.P., Yang, X.Y., Liu, L.F., et al.: Faster Bootstrapping with Multiple Addends.
Cryptology ePrint Archive, Report 2017/735 (2017). http://eprint.iacr.org/2017/735
9. Wang, W., Hu, Y., Chen, L., et al.: Accelerating fully homomorphic encryption using GPU.
In: High Performance Extreme Computing, pp. 1–5. IEEE (2013)
10. Dai, W., Doröz, Y., Sunar, B.: Accelerating NTRU based homomorphic encryption using
GPUs. In: IEEE High Performance Extreme Computing Conference, pp. 1–6. IEEE (2015)
11. Dai, W., Sunar, B.: cuHE: a homomorphic encryption accelerator library. In: International
Conference on Cryptography and Information Security in the Balkans, pp. 169–186.
Springer, Cham (2015)
12. Fast Fully Homomorphic Encryption Library over the Torus. https://github.com/tfhe/tfhe
13. Dai, W.: CUDA-accelerated Fully Homomorphic Encryption Library. https://github.com/
vernamlab/cuFHE
NADEEM: A Novel Reliable Data
Delivery Routing Protocol
for Underwater WSNs
Nadeem Javaid(B)
1 Introduction
In the last decade, UWSNs have gained a lot of attention from researchers
because of their potential to monitor underwater environment. UWSNs have
extensive variety of applications such as military defense, monitoring aquatic
environment, disaster prevention, oil/gas extraction, offshore exploration, com-
mercial and scientific purposes, etc. [1,2].
In underwater atmosphere protocols of terrestrial wireless sensor networks
(TWSNs) cannot work perfectly. The TWSN and UWSN have differences in
many aspects such as use of acoustic links instead of radio links, UWSNs topology
is more dynamic than TWSNs because nodes move freely with water currents
and change their position frequently, localization of nodes is difficult as compared
to TWSNs. In addition, UWSNs face many challenges like low bandwidth, high
propagation delay and high communication cost [3].
Various routing protocols are proposed to minimize the energy consumption
and avoid void occurrence [4,5]. To recover data from void node, fallback recovery
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 103–115, 2019.
https://doi.org/10.1007/978-3-030-15035-8_10
104 N. Javaid
of energy. For instance, node i transmits data to node j with energy E and when
j sends data to i, the battery dissipation remains the same. It only changes when
the distance between two nodes vary. (iii) All acoustic nodes has the power to
adjust communication range autonomously and sink node can receive multiple
packets at the same time without any data loss or collision. (iv) The effect
of water currents is in the direction of horizontal, while in vertical direction
the movement is almost negligible. In the following paragraph an overview of
Thorp’s propagation model [2] has been presented to model the energy con-
sumption according to the distance between the source and the destination.
The path loss due to unhindered propagation route for a signal having frequency
f over a distance l is given as [15]:
where s is the spreading factor and α(f ) is the absorption coefficient. The geom-
etry of propagation is described using the spreading factor k and its values
for spherical, cylindrical and practical spreading are: s = 2, s = 1, and s = 1.5
respectively. The absorption coefficient α(f ), in dB/km for f in KHz, is described
by the Thorp’s formula [2] as:
where Energyb and N oise0 are constants that reveal per bit energy transmis-
sion and noise power density on a non declining additive white gaussian noise
(AWGN) channel [3]. As in [16], rayleigh fading is used to model small scale
system where SNR has the following distribution probability:
∞
1 Y
ρl (Y ) = e− SN R(l) . (4)
0 SN R(l)
4 Proposed Work
In this section, we describe our proposed schemes in detail. Initially, we will dis-
cuss the beacon message dissemination in the network for configuration. Addi-
tionally, neighbor node selection mechanism will be presented for successful data
communication. Moreover, an efficient mate selection is discussed to show its
effectiveness for reliable data transmission.
The beacon dissemination is vital to enable the configuration of the nodes for
successful data communication. Once all nodes and sinks are deployed in the
network volume, collaboration is important to relay data in multihop fashion
from a source to destination. Thus, a beacon is transmitted by a sink node from
the surface of water which is also equipped with acoustic modem. This message
includes unique information like identification coordinates (x, y, z). The sinks
are equipped with positioning system and have the area facts of each other [2].
Similar is the case for sensor nodes. The beacon transmitted by sensor nodes con-
sists of sequence number, ID and x, y, z coordinates. The positioning system is
ineffective in underwater due to excessive frequency signal absorption [2]. There-
fore acoustic localization services are used for acquiring nodes information [6].
108 N. Javaid
In aforesaid equation, M (i) shows the mate selected for data communication
in the network. While i depicts the number of mates selected (i varies from 0
to N , where N could be any real positive integer). The right hand side of the
expression shows computation mechanism of priority factor. Here E(i) denotes
the energy of node i and it is multiplied with N number of neighbors to increase
its suitability because higher the neighbor number is, more will be backup avail-
able incase of transmission failure and 1 is added to avoid when N number is
0. The fundamental concept of this approach is to boost the data packet in the
direction of sinks on the surface. The selection of multiple neighbor nodes results
in less number of retransmissions of the data packets [18,20].
Moreover, one more factor is considered before nominating the forwarder
node, packet advancement (ADV) [18] to choose the forwarder node which leads
the packet towards the destination. The ADV is defined as the difference between
distance of a source node Si and destination node Di with difference of distance
between neighbor Y and Ni [2]. Where nj be the node which has a packet
to transmit and its neighbors set is represented as Nj (t) and set of sinks is
represented as Sj (t), where t is representing the time. Therefore, the neighbor
node set in geographic and opportunistic routing is given as:
Ni = {nk ∈ Nj (t) : ∃sv ∈ Sj (t)|Nd (nj , s∗i ) − Nd (nk , sv ) > 0}, (9)
where Si∗ is from set of sinks Si (t) and it is closest sink of node nj as:
here, ADV (nf ) = D(nj , s∗j ) − D(nf , s∗f ) is the nf ADV towards the closest sink,
s∗f ; lfj is the distance between source node nj and forwarder node nf and ρ(lfj , n)
is the probability of n bits over distance lfj and it is given in Eq. 7.
After the node set selection and next hop forwarder set selection assumes
fi ⊆ Ni is the set that is formed according to the priorities of NADV [2,20].
As the goal of geographic and opportunistic routing is to find the fi ⊆ Ni to
maximize the expected packet advance (EPA). The EPA of fi is formulated
in Eq. 11.
k
EP A(fi ) = Σd=1 N ADV (nd )Πj=0d−1
(1 − ρ(lij , n)). (12)
Ultimately, the set with maximum EPA is chosen as next forwarder node set.
The node with highest precedence is chosen from the forwarder set as a next hop
forwarder. It declares its location records to its neighbor node. If a node is not
decided as a next hop forwarder node for a longer period, its packet is discarded.
The TA-NADEEM has the same steps of NADEEM except the adjustment of
transmission to avoid void hole problem which is illustrated in Fig. 3. During
communication TA-NADEEM computes the number of neighbors of the for-
warder node after adjusting the transmission range. Let us assume a scenario as
depicted in Fig. 3, where node S has to send a data packet towards the destina-
tion, it has no neighbor in its transmission range. Instead of opting for fallback
procedure, it simply adjusts the transmission range and delivers the data packet
directly to the immediate node comes into its adjusted communication range.
transmission range incase of void occurrence to avoid packet drop. The imme-
diate forwarder is n1, S acquires the information about the neighbors of n1
to ensure that selected node is not a void. Thus, it looks into the neighbor
table of n1 and finds that it has two neighbors at higher depth than n1. Thus,
n1 is selected as the effective forwarder node to continue with the process of
data routing. If there is no neighbor in the communication vicinity of n1, then
further adjustment in transmission range is done until a suitable forwarder is
located. Additionally, it updates the information in packet queue if there is no
report regarding the packet delivery. It helps in avoiding redundant data packets
transmission resulting in less energy consumption which prolongs the network
lifespan. Moreover, in future, nodes look into the neighbor table and avoids the
void occurrence. In this way extra energy consumed in transmission adjustment
is accommodated.
Figure 4 illustrates the ratio of void nodes of NADEEM and its variants against
the GEDAR. It is evident from the results that void ratio decreases with the
increase in node number. The highest transmission failure ratio can be observed
of FA-NADEEM which falls significantly up to 250 nodes. Then again, after the
node density of 300, the amount of void node occurrence decreases to a certain
level, which remains consistent till 450 nodes. The decrease in ratio is because
of the fall back mechanism of FA-NADEEM which looks for an alternative route
112 N. Javaid
from all nodes in the communication range to recover and deliver data packet
successfully at the destination.
NADEEM has less void appearances at the end of the node increment because
it does not have transmission adjustment, fall back and depth adjustment mech-
anisms (Fig. 4). This help in saving node battery ultimately leading to less nodes
and more number of alive nodes. Thus, when the sparsity of the nodes decreases,
the probability of void occurrence also decreases.
GEDAR and TA-NADEEM show almost identical trend in rise and fall of
void proportion during the nodes communication. Although, the values are dif-
ferent and at the end TA-NADEEM beats the baseline scheme because of direct
transmission adjustment, instead of exchanging beacon message to find out the
different route (Fig. 4). When different route is not available then depth of the
node is adjusted, these aspects consume a lot of energy and suddenly node bat-
tery depletes. Whereas, TA-NADEEM adaptively adjusts its transmission power
and bypass the void node successfully.
From 150–200 nodes GEDAR shows sudden decrease in fraction of void nodes
(Fig. 4). As GEDAR considers movement of void node towards its neighbor node
without checking whether the neighbor nodes have enough forwarder nodes or
not, therefore it has high fraction of void nodes compared to proposed schemes.
0.5
NADEEM
0.45 GEDAR
FA-NADEEM
0.4
Fraction of local maximum nodes
TA-NADEEM
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
150 200 250 300 350 400 450
Number of nodes
Figure 5 shows that the energy tax gradually decreases with the increase in
density of network. The highest PDR is achieved by FA-NADEEM due to its
ability of finding data route from an alternate node within the communication
vicinity. When a trap of void node occurs, instead of dropping the packet, the
source node proceed with fall back procedure and looks in 360◦ to resume the
data transmission process.
NADEEM depicts moderate success ratio of packets transmission compared
to other schemes in Fig. 5. The performance is neither high nor low, it operates
NADEEM: A Novel Reliable Data Delivery Routing Protocol 113
0.25
Packet delivery ratio
0.2
0.15
0.1
0.05
0
150 200 250 300 350 400 450
Number of nodes
1
NADEEM
0.9 GEDAR
FA-NADEEM
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
150 200 250 300 350 400 450
Number of nodes
6 Conclusion
In this paper, we have proposed NADEEM, FA-NADEEM and TA-NADEEM
routing protocols to enhance the energy efficiency and minimize the fraction
of void node occurrence. The NADEEM minimizes the energy consumption by
avoiding immutable forwarder selection. It also achieves higher packet delivery
ratio. Whereas, FA-NADEEM avoids successfully the void node through send-
ing data from the alternate route. However, the use of alternate path increases
the number of hops which resulted in more energy consumption. Moreover, it
has higher throughput as compared to NADEEM. On the other hand, when
FA-NADEEM fails to find an another forwarding path then transmission of
NADEEM is adjusted to ensure data delivery at the destination node. Addition-
ally, TA-NADEEM looked for an immediate forwarder node and its neighbors to
ensure that selected node is not void. It has low network throughput, however,
minimum energy consumption for one packet. The results are validated through
extensive simulations. The outcome clearly shows that proposed schemes outper-
formed the baseline existing scheme (GEDAR) in terms of energy consumption
and void avoidance.
References
1. Javaid, N., Jafri, M.R., Ahmed, S., Jamil, M., Khan, Z.A., Qasim, U., Al-Saleh,
S.S.: Delay-sensitive routing schemes for underwater acoustic sensor networks. Int.
J. Distrib. Sens. Netw. 11(3), 532676 (2015)
2. Coutinho, R.W.L., Boukerche, A., Vieira, L.F.M., Loureiro, A.A.F.: Geographic
and opportunistic routing for underwater sensor networks. IEEE Trans. Comput.
65(2), 548–561 (2016)
3. Noh, Y., Lee, U., Lee, S., Wang, P., Vieira, L.F.M., Cui, J.-H., Gerla, M., Kim,
K.: Hydrocast: pressure routing for underwater sensor networks. IEEE Trans. Veh.
Technol. 65(1), 333–347 (2016)
4. Kheirabadi, M.T., Mohamad, M.M.: Greedy routing in underwater acoustic sensor
networks: a survey. Int. J. Distrib. Sens. Netw. 9(7), 701834 (2013)
NADEEM: A Novel Reliable Data Delivery Routing Protocol 115
5. Han, G., Jiang, J., Bao, N., Wan, L., Guizani, M.: Routing protocols for underwater
wireless sensor networks. IEEE Commun. Mag. 53(11), 72–78 (2015)
6. Coutinho, R.W.L., Boukerche, A., Vieira, L.F.M., Loureiro, A.A.F.: GEDAR: geo-
graphic and opportunistic routing protocol with depth adjustment for mobile
underwater sensor networks. In: 2014 IEEE International Conference on Com-
munications (ICC), pp. 251–256. IEEE (2014)
7. Yan, H., Shi, Z.J., Cui, J.: DBR: depth-based routing for underwater sensor net-
works. In: NETWORKING 2008 Ad Hoc and Sensor Networks, Wireless Networks,
Next Generation Internet, pp. 72–86 (2008)
8. Javaid, N., Jafri, M.R., Khan, Z.A., Qasim, U., Alghamdi, T.A., Ali, M.: iAMCTD:
improved adaptive mobility of courier nodes in threshold-optimized DBR protocol
for underwater wireless sensor networks. Int. J. Distrib. Sens. Netw. 10(11), 213012
(2014)
9. Azam, I., Javaid, N., Ahmad, A., Abdul, W., Almogren, A., Alamri, A.: Balanced
load distribution with energy hole avoidance in underwater WSNs. IEEE Access
5, 15206–15221 (2017)
10. Yu, H., Yao, N., Wang, T., Li, G., Gao, Z., Tan, G.: WDFAD-DBR: weighting
depth and forwarding area division DBR routing protocol for UASNs. Ad Hoc
Netw. 37, 256–282 (2016)
11. Ahmad, A., Javaid, N., Khan, Z.A., Qasim, U., Alghamdi, T.A.: (ACH)2 : routing
scheme to maximize lifetime and throughput of wireless sensor networks. IEEE
Sens. J. 14(10), 3516–3532 (2014)
12. Abbas, M.Z., Bakar, K.A., Ayaz, M., Mohamed, M.H., Tariq, M.: Hop-by-hop
dynamic addressing based routing protocol for monitoring of long range underwater
pipeline. KSII Trans. Internet Inf. Syst. (TIIS) 11(2), 731–763 (2017)
13. Yu, S., Liu, S., Jiang, P.: A high-efficiency uneven cluster deployment algorithm
based on network layered for event coverage in UWSNs. Sensors 16(12), 2103
(2016)
14. Javaid, N., Rasheed, M.B., Imran, M., Guizani, M., Khan, Z.A., Alghamdi, T.A.,
Ilahi, M.: An energy-efficient distributed clustering algorithm for heterogeneous
WSNs. EURASIP J. Wirel. Commun. Network. 2015(1), 151 (2015)
15. Heidemann, J., Stojanovic, M., Zorzi, M.: Underwater sensor networks: applica-
tions, advances and challenges. Phil. Trans. R. Soc. A 370(1958), 158–175 (2012)
16. Noh, Y., Lee, U., Han, S., Wang, P., Torres, D., Kim, J., Gerla, M.: DOTS: a prop-
agation delay-aware opportunistic MAC protocol for mobile underwater networks.
IEEE Trans. Mob. Comput. 13(4), 766–782 (2014)
17. Stojanovic, M., Preisig, J.: Underwater acoustic communication channels: propa-
gation models and statistical characterization. IEEE Commun. Mag. 47(1), 84–89
(2009)
18. Yang, H., Liu, B., Ren, F., Wen, H., Lin, C.: Optimization of energy efficient trans-
mission in underwater sensor networks. In: Global Telecommunications Conference,
GLOBECOM 2009, pp. 1–6. IEEE (2009)
19. Climent, S., Sanchez, A., Capella, J.V., Meratnia, N., Serrano, J.J.: Underwater
acoustic wireless sensor networks: advances and future trends in physical, MAC
and routing layers. Sensors 14, 795–833 (2014)
20. Noh, Y., Lee, U., Wang, P., Choi, B.S.C., Gerla, M.: VAPR: void-aware pressure
routing for underwater sensor networks. IEEE Trans. Mob. Comput. 12(5), 895–
908 (2013)
21. Yadav, S., Kumar, V.: Optimal clustering in underwater wireless sensor net-
works: acoustic, EM and FSO communication compliant technique. IEEE Access
5, 12761–12776 (2017)
An Architecture for Distributed
Ledger-Based M2M Auditing for Electric
Autonomous Vehicles
1 Introduction
Vehicles, once considered mechanical boxes and used as a primary source of com-
muting, are now transforming into smart cars. With the massive introduction of
software, modern cars are equipped with embedded sensors, actuators and com-
putational capabilities, including communication and storage. This surge in the
computation, communication, and automotive technologies gave rise to Intelli-
gent Transportation System (ITS) realized through Vehicular Ad hoc NEtworks
(VANET) [13]. Simultaneously, Electric Propulsion Vehicle (EPVs or EVs) were
introduced to address the drastic increase of global greenhouse gas (GHG) emis-
sions. Despite the numerous regulations that were made to address the ecological
problem, technological solutions are also necessary to be used in synergy. If we
also consider recent advances in the field of Artificial Intelligence (AI) we have all
the instruments to potentially reduce human error in transportation systems by
building self-driving vehicles. They are capable of autonomously steering, nav-
igating, vehicle agile maneuvering, making decisions, foreseeing potential acci-
dents and acting better than human drivers in some critical situations [14,31].
Having such machines at our disposal could result in numerous benefits, includ-
ing, but not limited to, fewer accidents, reduced traffic congestions, and enable-
ment for so-called Mobility-as-a-Service (MaaS) [16]. Let’s consider these topics
in details.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 116–128, 2019.
https://doi.org/10.1007/978-3-030-15035-8_11
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 117
Accidents. In the United States, motor vehicle crashes are one of the leading
causes of death, as approximately two million collisions happen every year [1,2].
A system which would eliminate the human error in the vehicle collisions would
result in a smaller number of vehicle crashes.
Traffic Congestion. It is the cause of numerous negative effects including
delays, air pollution and increased chance of collisions. This problem has not
been solved yet and is one of the main concerns for big cities and megalopolises.
Research has shown that only 5% of Autonomous Vehicles on our roads can
significantly reduce the stop-and-go traffic waves that can arise due to lane
changing [27].
MaaS. Mobility-as-a-Service is a concept that delivers users’ transportation
needs through a service provider (SP) [16]. The core idea is to have a fleet of
Self-driving vehicles (SDVs) which can perform transportation services for their
users. This could lead to the decline in car ownership and private car usage and
the expansion of the Sharing economy. Companies like Volkswagen are very inter-
ested in this type of transportation systems, as it has the potential to change the
automobile industry [17]. With the increasing popularity of EV, the need of hav-
ing a sufficient charging infrastructure appears. Currently, the most widespread
option is the use of Charging Stations (CSs) to deliver electrical energy from the
power source to the EV, ensuring the appropriate and safe flow of electricity.
CSs are the main interface between the vehicle, the utility and the user [6]. This
means that the service providers have enabled users to charge their vehicle by
plugging it into the CS, and then paying for the services provided through the
use of credit/debit cards. Although this may seem convenient for human use,
they are not suitable for the Machine-to-Machine (M2M) economy [3]. The vehi-
cles of the future will be driverless, requiring the absence of human interaction
in the charging process (this is a typical scenario for MaaS). In addition, such an
approach leads to fees imposed often by trusted third-party authorities. Finally,
they violate the bidirectional auditability and are not privacy-aware [15]. A new
system which would solve these issues is desirable.
DLTs and the Tangle. A new decentralized architecture and distributed com-
puting paradigm called the Blockchain has gained significant popularity. It rep-
resents an immutable ledger system implemented in a distributed way which
eliminates the need for a trusted third-party [19]. By using technologies relying
on distributed ledgers we could increase the transparency of the whole system
while maintaining the high level of its security. The research of DLT-related use
cases has been actively conducted for multi-agent systems [7,18], and one par-
ticular area which could be disrupted by the use of this innovative technology
is Internet-of-Things (IoT). The main reasons why IoT could benefit from the
DLT are M2M value transactions, increased security and the automation of pro-
cesses. However, technical limitations of the current blockchain architecture do
not allow it to be broadly used in this industry, primarily due to high transac-
tion fees and scalability issues [24]. Machines will need to make micropayments
between themselves, but in order to make them feasible, transaction fees should
be very minimal, preferably non-existent. In addition, transaction confirmation
118 D. Strugar et al.
times should be very short for the network to be scalable and ready to accept new
transactions. Recently, the Tangle has been introduced [21]. It is the next step
in the evolution of Distributed Ledger Technologies and is an underlying tech-
nology of the IOTA cryptocurrency. Based on a concept of a Directed Acyclic
Graph (DAG), Tangle still serves as a distributed database, but it has three
main advantages: it is scalable, involves no fees and allows for offline transac-
tions. These characteristics of the Tangle and its cryptocurrency make it possible
to become the backbone of the emerging Machine Economy.
With these concepts in mind, the cryptocurrency which uses the Tangle,
IOTA, aims to become the backbone of the emerging Machine Economy, by
allowing machines to trade resources (electricity in our case) and services with
each other without the involvement of the intermediary.
Outline of the Paper. This paper is focusing on the problem of charging Elec-
tric Autonomous Vehicles without human mediation using Distributed Ledger
Technologies. The problem is open and its solution may have significant impact
in shaping the society of the future. In this Section we have presented the general
scenario of EAV and Distributed Ledger Technologies, with particular attention
to IOTA, the cryptocurrency we are proposing for the payment method. In
Sect. 2 we present the overall Architecture with discussion of the technological
stack while in Sect. 3 we show the paradigmatic use case and what features the
platform should exhibit. Section 4 introduces the technological complexities of
the project, Sect. 5 discusses how to protect EAVs and CSs from malicious cheat-
ing in serving and billing, and Sect. 6 brings out some reflections and plans for
the future. In this paper we define the vision and the architecture of our frame-
work. The formal design of automotive system that can be applied to this case,
has been described in [10].
2 Platform Architecture
In this section we describe the overall system architecture and its organization
as well as the proposed technologies. Figure 2 depicts the architecture consisting
of three major layers:
1. Physical
2. Network
3. Services
This layer encapsulates all the hardware components embedded in the CS used
for sensing and gathering information about the charging process, as well as
communicating with external entities. The core parts of this layer are:
• kWh meter
• Main Controller (MC)
• EVSE Controller (EVSEC)
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 119
The main sensor in the CS is the kWh meter as it measures the amount
of electricity consumed by the EAV. Along with the current electricity market
price, this data would be constantly used to calculate the total sum that the
vehicle would pay to the service provider. The benefit of such an approach is
that users get the exact amount of electricity they have paid for.
Moreover, knowing the current electricity rate could help us in adjusting the
maximum power the EAV gets charged with, especially in the case of multiple
CS using the same power source. The device used to perform such an operation is
the Smart Electric Vehicle Supply Equipment (EVSE) [26], previously referred
to as EVSEC. This could deliver electricity in a controlled (smart) way, and
contribute to the development of the Smart Grid [4,25].
Lastly, we use MC to transfer the information gathered by components of
the Physical layer to higher levels of the architecture through the use of a com-
munication protocol specifically designed for IoT, Message Queuing Telemetry
Transport (MQTT).
online, depending on the protocol used. Since our architecture involves constant
M2M communication with subscription behavior, we consider MQTT connectiv-
ity protocol to be the most appropriate choice for inter-device communication
[12]. In our case, MQTT would ensure rapid transfer of both data and value with
high throughput.
In order to allow value transactions in such a system, a more secure solution
is needed. Recently, a bi-directional payment channel has been introduced which
meets the requirements stated above and conforms to the MQTT protocol by
nature. It is called the Flash Channel (FC) [9], and it allows both resource and
service providers to get paid for their services while they are performing them.
FCs enable real-time streaming of transactions, and allow machines to branch-
off the main network, transmit data and value via whichever protocol they find
suitable (MQTT in our case), and then merge the results with the main IOTA
Tangle. This means that there is no need to wait for transactions to be approved
by the IOTA network. By using it, only the starting and ending transactions of
the Flash channel will be published to the main network.
Once machines have finished transacting in FC, the final transaction, denoted
as tf , becomes ready to be published on the Tangle. When this happens, tf has to
verify two not yet approved transactions. It does so by performing computational
work which is used to verify if the transaction is trusted or not. This process is
called the Proof-of-Work (PoW). Similarly, other transactions tk =f , tl=k =f will
find and verify tf . In order for any transaction tx =f to be approved in the Tangle
network, two incoming transactions ty and tz have to verify it using their PoW.
Such an approach removes the need for miners, and therefore enables feeless trans-
actions. In addition, the network benefits from new transactions as they will verify
two unapproved ones and thus improve the scalability of the system [21].
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 121
Through the use of the Tangle, FCs and MQTT protocol for message transfer,
the Network Layer would be able to transfer sensor data from the Physical Layer
to the Services Layer rapidly and securely.
4 Research Challenges
The capabilities of IOTA are unprecedented and it is poised to revolutionize the
current micro-economy; however, in the domain of connected and autonomous
cars industry, there are challenges that need to be addressed. In this section, we
focus on the general IOTA-related and EAV-specific challenges.
Adaptation. IOTA is still in its infancy and at the moment, it is struggling with
growing the Tangle. Nevertheless, IOTA undoubtedly provides many benefits as
opposed to blockchain, for instance no transaction fees; however, the adaption
of IOTA among businesses is still a challenge and needs to address different
issues such as best practices of cryptography (the absence of which may lead
to security loopholes), conflict resolutions, approval of new members, and so
on. It is, at least at the moment, premature to guarantee its adoption among
the consumers and the businesses. In other words, it is still speculative what
the future will bring for IOTA, but the signs are convincing that IOTA will
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 123
to excel, it has to consider these challenges before they decide to remove the
coordinator from the network.
5 Threat Model
This section discusses how the proposed system would protect EAVs and CSs
from malicious cheating during the settlement of payments. One key characteris-
tic that makes the system we proposed in this article different is that it effectively
solves the Guest-Host problem (GHP) This problem presents the case of a guest
who wants to use a hotel room for some number of days and does not want to
get billed for the service used it. On the other hand, the host should not get paid
less than determined and agreed by both parties. In our case, we consider EAV
as the guest, and CS as the host of the GHP problem.
We consider three main scenarios: (A) Guest tries to get more charge than
it has paid for, (B) Host tries to bill the user more than it is agreed, and (C)
Either side tries to benefit from the interruption of the charging process.
A. Guest tries to get more charge than he/she has paid for
Problem. The most common scenario occurs when the guest gets the service as
promised, but does not want to pay the settled amount of money. For this very
reason, state-of-the-art electric CSs mostly require users to have special service
cards tied to their bank accounts and often passports/national IDs2 . This way
the CS can ensure that the user will get billed. Thus, scenarios where EAVs first
get the charge and then pay, are not convenient for service providers and are
seldom used.
Solution. Our approach does not assume that someone should pay first. Instead,
we look at the process of charging and billing as token streaming. Key idea is
to have a system where both charging and billing are happening at the same
time. That is, while EAV is getting the charge, it is paying for these services.
Our solution leverages IOTA, and their bidirectional Flash Channels (FCs) for
this specific use case. As we have discussed, FCs enable instantaneous, high-
throughput transactions. This fact is crucial in our scenario. Potentially, to be
entirely sure that the system is fair, the number of transactions per second
can be set. It would ensure that the users get exact amount of charge for the
amount they pay. Additionally, IOTA, by its nature, is privacy-aware, as IOTA
transactions do not include any personal information3 .
Thus, the guest (EAV) can not get more charge than it has paid for. And
additionally, the whole process of billing has numerous benefits, including, but
not limited to: trust, fairness for both parties, transparency and privacy.
2
But it is a prime example of Privacy violation - why service providers are able to
see users’ detailed private information?
3
IOTA transaction contains: transaction hash, signature message fragment, address
of the recipient, value, timestamp, bundle, nonce, and a few other properties that
reveal nothing about the identity of transacting parties.te.
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 125
Problem. While the state-of-the-art approach does in fact solve one aspect of
the GHP (host does get paid), it can be easily seen that this approach is not
perfect for the guest. First, EAV exposes its private information, then pays for
the service, and in the end, has to trust the CS that it will deliver the promised
charging services. And not only that, CS can hide the current market price, and
bill users way more than they should be billed. Therefore, users must be able to
see the price they are getting the electricity for.
Solution. Our proposed scheme solves this problem very effectively. At each
point in time, users are billed with the amount of money which corresponds to
the amount of electricity they get, as there are no fees. The whole process of
charging and billing is transparent for the guest of the GHP. If the host tries
to bill the guest more than it is agreed, that malicious activity can be easily
identified, and the host could have reporting contract violation warning.
Problem. One notable case occurs when either the guest or the host decides to
quit the charging process by force. In traditional charging station systems, there
are very few ways to deal with such a situation. What service providers do, is
employ people to regulate the whole process that degrades the M2M concept.
Solution. As in the previous two cases, payment channels handle also provide a
very elegant solution. Either side could terminate the charging process by force,
and the only consequence would be that the channel closes. When that happens,
CS constructs a bundle that takes the balance of the channel, and divides it
amongst the channel’s users accordingly. These bundles are then attached to the
network, and neither side benefits from this interruption.
6 Conclusions
The emergence of Intelligent Technologies is pervading every moment of our
life. While cars themselves may externally look similar to decades ago (apart
from design trend consideration), a dramatic revolution is happening inside the
cabin. This revolution could be described by the single word: software. Soft-
ware is becoming dominant, at times pervasive, of everyday existence. This cer-
tainly demands for high attention on privacy and ethical issues, however expands
our possibilities beyond the unthinkable. The same way that we have intelli-
gent portable devices, we have intelligent buildings [22,23] that can significantly
improve our lives or the lives of people with impaired capacities [20], and so we
are aiming at having intelligent cars. In this paper we discussed the synergy of
three fundamental technologies: Autonomous Vehicles, Electric Cars and cryp-
tocurrencies. What we propose is their cooperation in order to implement M2M
126 D. Strugar et al.
charging with realization of a billing framework for EAVs via IOTA technol-
ogy. Future work will involve the full realization of the vision here described, at
first with a computer graphical simulation, and finally with a full scale develop-
ment. After these three phases have been completed, we are going to present the
project to authorities in order to upgrade the project to a larger scale. Further
refinements could include a microservice-based architecture, aimed at increasing
scalability [8], which is already part of our research agenda. At a larger scale, the
vision could be expanded foreseeing a marketplace of individuals and companies
selling out energy in excess or energy generated precisely for this purpose. Once
the technology will be in place, this would create a specific market run on the
idea of cryptocurrency as a value-exchanger in a scenario where cars can drive
and stop for charge in a completely autonomous manner.
References
1. Critical reasons for crashes investigated in the national motor vehicle crash causa-
tion survey. Technical report, U.S. Department of Transportation, 1200 New Jersey
Ave, SE, Washington, DC, 20590, USA, February 2015. https://crashstats.nhtsa.
dot.gov/Api/Public/ViewPublication/812115
2. Traffic safety facts, 2015 - a compilation of motor vehicle crash data from the
fatality analysis reporting system and the general estimates system. Technical
report, U.S. Department of Transportation, 1200 New Jersey Ave, SE, Wash-
ington, DC, 20590, USA (2015). https://crashstats.nhtsa.dot.gov/Api/Public/
ViewPublication/812384
3. Bojkovic, Z., Bakmaz, B., Bakmaz, M.: Machine to machine communication archi-
tecture as an enabling paradigm of embedded internet evolution. In: International
Conference on ACE, pp. 40–45 (2014)
4. van den Brink, H.: How ElaadNL built a PoC charge station running fully on IOTA,
and IOTA only (2017). https://medium.com/@harmvandenbrink/how-elaadnl-
built-a-poc-charge-station-running-fully-on-iota-and-iota-only-e16ed4c4d4d5
5. Cheng, K.W.E., Divakar, B.P., Wu, H., Ding, K., Ho, H.F.: Battery-management
system (BMS) and SOC development for electrical vehicles. IEEE Trans. Veh.
Technol. 60(1), 76–88 (2011)
6. Cities, C.: Plug-in electric vehicle handbook for public charging station hosts. US
Department of Energy Publication No. DOE/GO-102012-3275 (2012)
7. Danilov, K., Rezin, R., Kolotov, A., Afanasyev, I.: Towards blockchain-based robo-
nomics: autonomous agents behavior validation. In: 9th International Conference
on Intelligent Systems. IEEE (2018)
8. Dragoni, N., Lanese, I., Larsen, S.T., Mazzara, M., Mustafin, R., Safina, L.:
Microservices: how to make your application scale. In: 1st Convergent Cognitive
Information Technologies, pp. 95–104 (2017)
9. Freiberg, L.: Instant & feeless flash channels (2017). https://blog.iota.org/instant-
feeless-flash-channels-88572d9a4385
10. Gmehlich, R., Grau, K., Iliasov, A., Jackson, M., Loesch, F., Mazzara, M.: Towards
a formalism-based toolkit for automotive applications. In: 1st FME Workshop on
Formal Methods in Software Engineering (FormaliSE) (2013)
11. Guardian: How world’s first electrified road charges moving vehicles (2018). www.
youtube.com/watch?time continue=56&v=VZNHZnyxCm8
An Architecture for Distributed Ledger-Based M2M Auditing for EAVs 127
30. Yuan, Y., Wang, F.Y.: Towards blockchain-based intelligent transportation sys-
tems. In: IEEE 19th International Conference on Intelligent Transportation Sys-
tems (ITSC), pp. 2663–2668. IEEE (2016)
31. Zubov, I., Afanasyev, I., Shimchik, I., Mustafin, R., Gabdullin, A.: Autonomous
drifting control in 3D car racing simulator. In: 9th IEEE International Conference
on Intelligent Systems. IEEE (2018)
Is QUIC Quicker Than TCP?
A Performance Evaluation
1 Introduction
The internet since it’s invention has become the most efficient and effective ways
to communicate. Nowadays we depend on the internet for a myriad of things,
including business, online transactions, education, inventory, marketing, banking
and communication etc. It is important that we ensure a fair, smooth and efficient
internet all over the globe. However, since online traffic is constantly changing
at any given time or place, it is difficult to ensure the same.
There is a need for globally accepted internet protocols that will ensure effi-
cient and fair use of internet between any two machines. TCP [9] is one such
protocol that has been around 1982. TCP is a widely used protocol that ensures
fair and efficient network with built in effective congestion control mechanisms
for traditional web applications. However, with increase in number of users, web
traffic, content size and interactive web applications, TCP has failed to maintain
optimal network conditions.
QUIC [8] is a protocol developed by Google to address many of the issues
encountered with TCP. As shown in Fig. 1, it is an application layer protocol
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 129–138, 2019.
https://doi.org/10.1007/978-3-030-15035-8_12
130 Saurabh et al.
that runs on the top of UDP [10] protocol. With QUIC, Google has tried to
provide some key features including reduced connection establishment time in
0-RTT(average), better connection migration, improved congestion control and
reduction of HOL blocking. Google is still rethinking some design decisions like
error recovery features as previously implemented FEC didn’t prove to be too
useful. QUIC implements TCP like properties in the application layer rather
than in the transport layer.
Google chrome uses QUIC by default and all the Google servers employ this
protocol. Referring [2] since its introduction in 2013, traffic using this protocol
increased significantly, currently approximately 58% mobile traffic and 64% video
traffic uses QUIC (Fig. 2). Many of the Google owned web applications like
youtube, gmail, drive etc. have already started using QUIC for their internet
traffic. However, it is still in experimental stage and there is a lot of room for
improvement. A dummy QUIC server and QUIC client, only able to be used for
integration testing and therefore not scalable has been released publicly online
for anyone to use and toy with [6].
1.1 Contribution
Our paper explores QUIC and TCP’s performance over various control param-
eters, and analyzing the raw data to compare them. Section 2 discusses related
work. Section 3 gives a detailed discussion of TCP and QUIC. Section 4 talks
more about the experiment setup and methodology used to achieve results in this
paper. Section 5 discuss the performance analysis and results obtained. Section 6
presents conclusions based on these results and finally Sect. 7 discuss future work.
2 Related Work
There has been no significant work done over performance of QUIC over con-
gested networks under variable control parameters. Megyesi et al. [12] focuses
on page load times over QUIC as compared to TCP + HTTP1.1 under lab con-
ditions, Carlucci et al. [3] also analyses QUIC in the terms of page load times.
Hussein et al. [1] observed QUIC performance over 3D web. Quentin et al. [5]
compares performance of MPTCP with their own implementation of multipath
QUIC under lossy links. Lychev et al. [11] prioritizes cryptographic schemes
present in QUIC. Sarah et al. [4] focuses on page load times under QUIC with
added packet loss rates.
On paper QUIC looks very powerful and successful in improving on TCP’s
limitations, but it is still an immature technology carrying only limited func-
tionality unlike TCP which can be used for file sharing, remote shell access and
network routing. QUIC might still be too far to being a replacement for TCP, but
given enough time we may see many of its successful features being implemented
in TCP.
3 Background
This section briefs built-in mechanisms provided with both TCP and QUIC.
This section uses TCP Reno as reference to briefly describe various built-in
concepts of the TCP protocol.
132 Saurabh et al.
Slow Start. TCP ensures that no new connections steal bandwidth away from
existing connections. This is done by exponentially increasing the size of sliding
window.
Fast Recovery. This improvement was added to TCP Reno and called New
Reno when it was discovered that most fast re-transmit actions occur during
mild congestion. Instead of using slow start TCP increases congestion window
size for every duplicate ACK received after the third one until we encounter a
non-duplicate ACK. The congestion window size and slow start threshold are
then reduced to reflect packet loss.
Is QUIC Quicker Than TCP? 133
Forward Error Correction (FEC). QUIC can recover from lost packets
without having to wait for retransmission of this packet using FEC packets.
QUIC compliments a group of packets with a FEC packet which contains parity
of packets in the FEC group. The sender may decide beforehand whether to use
FEC or not to optimize link conditions.
4 Methodology
For TCP, this paper uses an apache2 server for server side whereas wget is used
to download files over HTTP/2+TLS over the client side. For QUIC, quic server
and client are used which uses iproute2 and netem to send the file from server
to client. The architecture is shown in Fig. 5.
The various control parameters were varied over different values and for each
set of parameter values, the tests were run for a total of 5 times. The bandwidth
rates were 100 Mbps 40 Mbps and 5 Mbps while delay rates were 10 ms, 50 ms
and 100 ms, packet loss rates were 0%, 0.5%, 1.0% and 5%, finally either spikes
were present or they were absent. The results are then plotted using python class
‘matplotlib’.
5 Analysis
The project ran in 3 phases.
Clean the dirty data and extract only needed information like timestamp and
bytes in ‘./clean’ folder. Average the data results for same instances of test, on
average each test is run for 5 times by default.
5.3 Visualization
We generate plots for performance bandwidth, overhead and time vs delay, band-
width, packet loss.
The following observations were made from obtained results From Fig. 6
we observe, QUIC generally performed better than TCP when packet losses
were small, but for higher packet losses there was no significant difference in
throughput. We can interpret from Fig. 7 that, under high bandwidth and delay
QUIC will perform much better than TCP. From Fig. 12 it can be inferred that,
QUIC performs generally better than TCP but has to make a trade on overhead
Fig. 6. Throughput vs packet loss under 100 ms delay and 100 Mbps bandwidth
136 Saurabh et al.
Fig. 7. Throughput vs delay under 5% packet loss and 100 Mbps bandwidth
Fig. 8. Time series graph under 5 Mbps bandwidth 0% packet loss and 100 ms delay
Fig. 9. Time series graph under 100 Mbps Fig. 10. Throughput comparison of TCP
bandwidth 1% packet loss and 100 ms vs QUIC when jitter is present 1
delay
size. From Figs. 8 and 9 it can be seen that, QUIC is more stable TCP under
favourable conditions. From Figs. 10 and 11 we can infer that TCP may outper-
form QUIC if jitter is present. Finally from Figs. 13 and 14 we observe, QUIC is
very unfair when competing for bandwidth as opposed to TCP.
Is QUIC Quicker Than TCP? 137
Fig. 13. TCP fairness vs no. of flows Fig. 14. QUIC fairness vs no. of flows
6 Conclusions
QUIC performed similar to TCP when packet losses were involved, this is because
FEC has been currently disabled and needs to be explicitly specified in the header
to be used and therefore QUIC interprets packet losses the same way TCP does
and re transmits them.
It is evident that QUIC transfers happen faster than TCP transfers which can
be credited to high throughput and efficient bandwidth usage. QUIC is generally
more stable than TCP, this may be because of high NACK size of 256 for QUIC,
which can continue in-taking packets even when previous packets were dropped.
Even though QUIC uses same congestion control algorithm as TCP Cubic,
still QUIC almost always uses more than it’s fair share of bandwidth when
competing with TCP flows. This may be because of large congestion window
size of QUIC and it’s tendency to increase its window size more aggressively
than TCP, which allows it to use bandwidth available more greedily.
We observed that TCP can outperform QUIC when jitters is present, this
shows that QUIC is still incomplete and not perfect. We did not observe any
significant differences in QUIC and TCP even though theoretically QUIC has
way better stats. All in all these results should be taken with a grain of salt,
since they were performed at home over restricted network and not by Google
on their servers where they have millions of users to test on.
138 Saurabh et al.
QUIC is still lacking in many ways which is evident from Google continually
trying to improve and better it, particularly error recovery has a huge scope for
improvement since it is practically disabled right now.
7 Future Work
For future work it might be interesting to explore the following aspects. Ana-
lyzing stream multiplexing in QUIC and observing its effects on load times of
HTTP pages. Evaluation and comparison of TCP and QUIC during connection
migration. Performance evaluation of QUIC over a wireless network. Comparing
performance of QUIC and TCP with other protocols like MPTCP and SCTP.
Proposing schemes for further optimization of QUIC protocol.
References
1. Bakri, H., Allison, C., Miller, A., Oliver, I.: HTTP/2 and QUIC for virtual worlds
and the 3D web? Procedia Comput. Sci. 56, 242–251 (2015)
2. Bouzas, S.: Why the meteoric rise of google quic is worrying mobile opera-
tors. https://2nuzlzatrf-flywheel.netdna-ssl.com/uploads/Chart-for-25Jan18-blog.
jpg. Accessed 01 Dec 2018
3. Carlucci, G., De Cicco, L., Mascolo, S.: HTTP over UDP: an experimental investi-
gation of QUIC. In: Proceedings of the 30th Annual ACM Symposium on Applied
Computing, pp. 609–614. ACM (2015)
4. Cook, S., Mathieu, B., Truong, P., Hamchaoui, I.: QUIC: better for what and for
whom? In: IEEE International Conference on Communications (ICC 2017) (2017)
5. De Coninck, Q., Bonaventure, O.: Multipath QUIC: design and evaluation. In:
Proceedings of the 13th International Conference on emerging Networking Exper-
iments and Technologies, pp. 160–166. ACM (2017)
6. Google: Playing with QUIC. https://www.chromium.org/quic/playing-with-quic.
Accessed 01 Dec 2018
7. Jain, R.: Fairness: how to measure quantitatively? ATMF/94-0881 (1994)
8. Langley, A., Riddoch, A., Wilk, A., Vicente, A., Krasic, C., Zhang, D., Yang, F.,
Kouranov, F., Swett, I., Iyengar, J., et al.: The quic transport protocol: design and
internet-scale deployment. In: Proceedings of the Conference of the ACM Special
Interest Group on Data Communication, pp. 183–196. ACM (2017)
9. Lin, F., Li, X., Li, W.: Research on TCP protocol in wireless network and net-
work simulation. In: 4th International Conference on Wireless Communications,
Networking and Mobile Computing, WiCOM 2008, pp. 1–4. IEEE (2008)
10. Liu, X., Cheng, L., Bhargava, B., Zhao, Z.: Experimental study of TCP and UDP
protocols for future distributed databases (1995)
11. Lychev, R., Jero, S., Boldyreva, A., Nita-Rotaru, C.: How secure and quick is
quic? Provable security and performance analyses. In: 2015 IEEE Symposium on
Security and Privacy (SP), pp. 214–231. IEEE (2015)
12. Megyesi, P., Krämer, Z., Molnár, S.: How quick is QUIC? In: IEEE International
Conference on Communications (ICC), pp. 1–6. IEEE (2016)
PRAN: A Provenance Based Model
and Prototype to Strengthen Authentication
1 Introduction
2 Problem Statement
3 Related Work
Our research work is focused more on provenance based security, and its application in
the area of authentication. Over the last few years, in the provenance related literature,
there have been some works done in the context of PAC: preserving integrity and
confidentiality [5, 6].
Recently, there has been considerable work done on PBAC. It has been recognized
that dependency relationships captured in a provenance aware system (PAS) [13] can
be used to control access to data. The work in [5] designed a basic PBAC model that
supports dynamic separation of duty, work flow. There have been discussions on access
control languages for PBAC [12], provenance aware access control policies [7], and
extension of PBAC model in [8]. To the best of our knowledge, there is no work in the
literature that addresses authentication problems using provenance utilities.
In the literature, there are some works that addresses password authentication
problems. Password stealing is addressed by strengthening passwords: using graphical
user interfaces [20] on the premise that humans remember images more than textual
passwords; proposed strong password protocols, and safe password entry interfaces
[21]; user authentication protocols [22]. These works are limited in usability aspects
due to obtrusiveness.
The work in [11] proposes a frame work for selecting an active authentication
factor based on contextual factors to balance security and usability of authentication. It
addresses the intrusive aspects of MFA but does not deal with the improving a way to
confirm user’s identity. The technique NIAS [14] uses knowledge of user to understand
the user’s identity using probabilistic methods based on users’ past access. Our work
deals with detecting potential anomalies as a result of password sharing/stealing based
on authentication provenance.
The behavioral bio-metrics [15] comes close to our work because of unobtrusive
way of capturing keystrokes, mouse dynamics to verify the claim of a user. However,
behavioral bio-metrics is new, and a lot of work needs to be done to overcome accuracy
issues, and has open challenges [17]. Our work is different from behavioral bio-metrics
in that we capture information about entities that are involved in a user authentication,
and use that information for evaluating the current authentication context based on
authentication policies that are specified in terms of abstractions. Existing literature in
password solutions has not paid much attention to password sharing problem.
142 R. S. C. Reddy and S. R. Gopu
Let UA, SA, APA, ACA, CONA are a set of user, system, application, activity, and
contextual attributes such that uaɛUA, saɛSA, apaɛAPA, acaɛACA, and conaɛCONA
uniquely represented. Let AR is a set of authentication requests such that
Let P be a set of predicates that define relation between AG, EN, AC. Let PATTR be
a set of attribute type, value pairs applied to AG, EN, AC as a triple such that
8rl e RL, ua e UA, sa e SA,apa e APA, aca e ACA, cona e CONA, pabsePABS: rl !
ðua, sa, apa, aca, cona, pabs):
ð9Þ
5 Prototype
We have developed the prototype to demonstrate the PRAN model. We have used
Apache jena1, prov tool2, and WSO2 Balana3 XACML java libraries:Apache jena for
capturing, querying the authentication provenance; prov tool java libraries for showing
the graphical representation of the authentication provenance; XACML [18] for policy
specification and evaluation. The Fig. 3 demonstrates the PRAN prototype architecture.
1
https://jena.apache.org/.
2
http://lucmoreau.github.io/ProvToolbox/.
3
https://github.com/wso2/balana.
PRAN: A Provenance Based Model and Prototype to Strengthen Authentication 145
6 Prototype Evaluation
The PRAN checks whether there are consecutive login failures for the system MAC
of the present login request. Usually, in case of brute force attacks, requests may come
from different systems or same system. To prevent anomalous login requests from
different systems, authentication policy rule can be changed to include other system or
contextual attributes.
ð11Þ
where “current_location” and “current_mac” are the location, and system MAC of the
authentication context.
It may so happen that the user “Alice” carries her laptop and actually present in that
location and submitted the login request. In this scenario also, the system will allow
login because Alice system will match with the MAC used in the previous login.
PRAN: A Provenance Based Model and Prototype to Strengthen Authentication 147
Use case 2 - Alice and Bob Are Co-located—Bob Uses Alice’s Credentials
In this scenario, the PRAN prototype will deny the login request based on the
evaluation.
ð12Þ
Here, location matches with the previous login location. However, system MAC will
not match because Bob’s system MAC would be different from the Alice’s system
MAC.
Use Case 3 - Alice and Bob Are Co-located—Alice Uses Alice’s Credentials
This is a genuine scenario. However, the policy specifies “allow Alice to use Bob’s
system only if Alice’s system is crashed”.
ð13Þ
In this scenario, it checks that whether the user u logged into her system. In case of
system crash, the user u would not have logged into her system. Hence, the PRAN will
allow user login to the application using Bob’s system.
Use Case 4 - Bob Uses Alice’s Credentials on Alice’s System
This is a tricky scenario. The prototype cannot identify this scenario based on the
policy rules because Alice’s credentials are used on Alice’s system. However, the
administrator can specify an additional rule that if Alice has a habit of sharing cre-
dentials then inform the application to ask the user to provide additional information—
it is up to the application how it asks for additional information.
Use Case 5 - Eve Manages to Steal Alice’s Credentials
In this scenario, Eve manages to steal password by using shoulder surfing attack. Here,
we assume that Eve attempts to login to the application using a different system other
than Alice’s system. In this scenario, the rule the use case scenario 2 will detect the
anomaly.
ð14Þ
If current location matches with the previous login location and current system
MAC matches with the previous system login’s MAC or if location mismatches and
MAC matches with the previous login’s MAC.
148 R. S. C. Reddy and S. R. Gopu
Use Case 6 - Alice Uses Her Personal System to Login to the Application
In this scenario, Alice declares that she will use her home system which will be
captured as provenance. The check below will allow Alice to login to the system
ð15Þ
Even though, the PRAN prototype is promising in detecting anomalous login events, it
has a few limitations and challenges. Since it is based on provenance of different
entities, availability of all required attributes is critical—for instance, determining
location is possible only if, the user is willing to share location. The prototype needs
administrators to be knowledgeable about password sharing/stealing scenarios, and
provenance abstractions, to be able to create authentication policies for such scenarios.
There are some challenges to overcome in order to make the prototype a real-time
implementation:
• Collection—making provenance collection trustworthy, accountable, and seamless.
• Maintenance—dealing with performance overhead associated with processing,
retrieving and storing humongous provenance.
• Security—ensuring provenance security to prevent unauthorized modification.
PRAN: A Provenance Based Model and Prototype to Strengthen Authentication 149
We have presented the PRAN model, and the prototype. We evaluated the prototype
with use case scenarios, and provided its limitations and challenges. We observe that
this work is certainly promising to be a useful attempt in the area of strengthening
authentication leveraging provenance. We also observe that provenance abstraction
would need to be generated for all the scenarios utilizing authentication provenance
causal relationships to enable the administrators create policies. Even though, we
proposed the PRAN model for password based authentication, it can be used for other
types of authentications also. Our recommendations for future research are: utilizing
provenance as an authentication factor which can strengthen single-factor authentica-
tion, and can be unobtrusive, and convenient; dealing with performance overhead
associated with querying, processing of humongous provenance:is there any way, a
system can remove noise (unwanted provenance) from the collected provenance; uti-
lizing provenance for other security aspects with showing some advancement in
usability, convenience, scalable aspects.
References
1. Groth, P., Moreau, L.: PROV-overview. An overview of the PROV family of documents
(2013)
2. Braun, U., Shinnar, A.: A security model for provenance (2006)
3. Hasan, R., Sion, R., Winslett, M.: The case of the fake Picasso: preventing history forgery
with secure provenance. In: FAST, vol. 9, pp. 1–14 (2009)
4. Syalim, A., Nishide, T., Sakurai, K.: Preserving integrity and confidentiality of a directed
acyclic graph model of provenance. In: IFIP Annual Conference on Data and Applications
Security and Privacy, pp. 311–318. Springer, Heidelberg (2010)
5. Park, J., Nguyen, D., Sandhu, R.: A provenance-based access control model. In: 2012 Tenth
Annual International Conference on Privacy, Security and Trust (PST), pp. 137–144. IEEE
(2012)
6. Nguyen, D., Park, J., Sandhu, R.: Dependency path patterns as the foundation of access
control in provenance-aware systems. In: TaPP (2012)
7. Sun, L., Park, J., Sandhu, R.: Engineering access control policies for provenance-aware
systems. In: Proceedings of the Third ACM Conference on Data and Application Security
and Privacy, pp. 285–292. ACM (2013)
8. Nguyen, D., Park, J., Sandhu, R.: A provenance-based access control model for dynamic
separation of duties. In: 2013 Eleventh Annual International Conference on Privacy, Security
and Trust (PST), pp. 247–256. IEEE (2013)
9. Sun, L., Park, J., Nguyen, D., Sandhu, R.: A provenance-aware access control framework
with typed provenance. IEEE Trans. Dependable Secur. Comput. 13(4), 411–423 (2016)
10. Khan, H.Z.U., Zahid, H.: Comparative study of authentication techniques. Int. J. Video
Image Process. Netw. Secur. (IJVIPNS) 10(4), 9–13 (2010)
11. Bonneau, J., Herley, C., Van Oorschot, P.C., Stajano, F.: The quest to replace passwords: a
framework for comparative evaluation of web authentication schemes. In: 2012 IEEE
Symposium on Security and Privacy, pp. 553–567. IEEE (2012)
150 R. S. C. Reddy and S. R. Gopu
12. Cadenhead, T., Khadilkar, V., Kantarcioglu, M., Thuraisingham, B.: A language for
provenance access control. In: Proceedings of the First ACM Conference on Data and
Application Security and Privacy, pp. 133–144. ACM (2011)
13. Miles, S., Groth, P., Munroe, S., Moreau, L.: PrIMe: a methodology for developing
provenance-aware applications. ACM Trans. Softw. Eng. Methodol. (TOSEM) 20(3), 8
(2011)
14. Al-Karkhi, A., Al-Yasiri, A., Linge, N.: Discreet verification of user identity in pervasive
computing environments using a non-intrusive technique. Comput. Electr. Eng. 41, 102–114
(2015)
15. Yampolskiy, R.V., Govindaraju, V.: Behavioural biometrics: a survey and classification. Int.
J. Biom. 1(1), 81–113 (2008)
16. Prud, E., Seaborne, A.: SPARQL query language for RDF (2006)
17. Buriro, A.: Behavioral biometrics for smartphone user authentication. Ph.D. diss., University
of Trento (2017)
18. Extensible Access Control Markup Language (XACML), version 3.0. Oasis Standard, 22
January 2013
19. Carroll, J.J., Klyne, G.: Resource Description Framework ({RDF}): Concepts and Abstract
Syntax (2004)
20. Khan, W.Z., Aalsalem, M.Y., Xiang, Y.: A graphical password based system for small
mobile devices (2011). arXiv preprint: arXiv:1110.3844
21. Ruoti, S., Andersen, J., Seamons, K.E.: Strengthening password-based authentication. In:
Way@ Soups (2016)
22. Sun, H.-M., Chen, Y.-H., Lin, Y.-H.: oPass: a user authentication protocol resistant to
password stealing and password reuse attacks. IEEE Trans. Inf. Forensics Secur. 7(2),
651–663 (2012)
23. Reddy, R.S.C., Gopu, S.R.: A dynamic delegation model and prototype for an enterprise
context. In: 2017 IEEE Conference on Open Systems (ICOS), pp. 71–76. IEEE (2017)
The Deepest Vertex First Reboot:
Rebooting Network Edge Switches
in a Campus Network
1 Introduction
Even in the era of Software Defined Network (SDN) or Software Defined Infras-
tructure (SDI), network edge switches require to be rebooted in some situations,
e.g., updating a firmware, configuring a special behavior and so on. We, Tottori
University, replaced almost all network switches in our campus network with
newer network switches made by AlaxalA in September 2017. We are now facing
a serious problem that an edge switch requires a reboot to activate a new digi-
tal certificate for web authentication after installing the certificate. As a digital
certificate expires within few years or even few months, we have to periodically
install a new digital certificate into all edge switches, and reboot them. When an
edge switch is rebooted, an end host becomes disconnected, and this downtime
should be minimized.
In order to dig into the best order or timing to reboot an edge switch, this
paper proposes a method, the deepest vertex first reboot, to reboot edge switches.
This method represents a campus network as an unweighted directed rooted
spanning tree graph. This method then enables to obtain a set of edge switches
that can be simultaneously rebooted. This method then defines to reboot edge
switches in the deepest vertex first fashion, i.e., rebooting the farthest edge
switches from a core switch. This method can avoid unnecessary waiting time
for a switch reboot and then minimize the total downtime of a network.
Note that this paper focuses the case where almost all of edge switches require
reboots but non-reboot switches such as core, distribution and some edge switches
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 151–160, 2019.
https://doi.org/10.1007/978-3-030-15035-8_14
152 M. Ohmori et al.
do not. It is out of the scope of this paper to reboot core and distribution
switches.
The rest of this paper is organized as follows. Section 2 defines preconditions
of this paper, and presents a campus network representation in this paper in a
graph theory fashion. Section 2 then proposes the rebooting method, the deepest
vertex first reboot. Section 3 explains a prototype implementation of the proposed
rebooting procedure including how to obtain a campus network graph. Section 4
then evaluates the proposed method using the prototype implementation in an
actual campus network. Section 5 finally concludes this paper.
v0
v1 v2
v3 v4 v5 v6
v7 v8 v9
rooted spanning tree graph G. V (G) and E(G) then denote a set of vertices and
arcs, respectively. A switch is a vertex, and i-th switch is denoted by vi ∈ V (G).
Let v0 be a vertex of a 0-th switch, i.e., a root. Let v0 be the only source vertex
with indeg(v0 ) = 0 in G, i.e., a core switch, while there are sink vertices vi with
outdeg(vi ) = 0 which are connected to no downstream switch. Downstream here
is a direction from the root v0 to other vertices. On the other hand, upstream is a
direction toward a root v0 . An upstream switch vi is then adjacent (i.e., physically
and directly connected) to another downstream switch vj by an unweighted arc
vi vj ∈ E(G) where i < j. On the other hand, suppose the downstream switch
vj has no arc back to the upstream switch vi for simplicity in discussions and
computational efficiency. Note that traffic may actually go through from vj to
vi in an actual network. Suppose non-reboot switches may not require to be
rebooted due to some reasons such that they accommodate no end host or do
not implement web authentication. Non-reboot switches depicted in Fig. 1 are v0 ,
v1 and v5 , and they are marked by prime ( ).
Note that a root vertex v0 may also be able to be contracted in theory. A root
vertex v0 is, however, not contracted for simplicity in discussions in this paper.
Note that G/V (G) is then a spanning tree as well.
A depth of a vertex is the number of arcs from a root in a graph, and a depth
of a root is 0. Let us focus on a depth in a contracted graph G/V (G), not in
a graph G. In Fig. 2, the depths of {v0 }, {v2 , v3 , v4 } and {v6 , v7 , v8 , v9 } are 0, 1
and 2, respectively. A depth of each vertex can be obtained as follows.
Let A be an adjacency matrix for G/V (G), and aij be an element (i, j) of A
where i and j denote indices of vi and vj , respectively. Let anij be an element (i, j)
of An . Note that indices of contracted vertices are not included in A and An .
154 M. Ohmori et al.
v0
v3 v4 v2
v7 v8 v9 v6
Corollary 1. For an index of each vertex in G/V (G) except for a root v0 , there
exists one and only one n ∈ N such that an0j = 1, and n will be a depth of vj .
Proof. aij ∈ {0, 1} since G/V (G) is an unweighted graph. In accordance with
a nature of an adjacency matrix, if anij = 0 then n is the number of vertices
of a simple path from vi to vj and anij is the number of simple paths. Actually
anij ∈ {0, 1} because G/V (G) is an unweighted spanning tree, and there exists
only one simple path from vi to vj if exists. There also exists one and only one
simple path from v0 to vj . There, hence, exists one and only one n ∈ N such that
an0j = 1, and n is then a depth of a vertex vj .
Proof. A is an upper triangular matrix because G/V (G) is a directed graph and
there are no back arc to an upstream vertex. The main diagonal components aii
of A are all zeroes since an arc from vi to vi itself never exists. A is then a strictly
upper triangular matrix, and nilpotent, say with Ar = 0 [1]. A then holds {∃ m ∈
N | Am = O ∧ Am+1 = O}. The maximum depth of vertices is then m.
Proof. An = O where n > m. Since G/V (G) is a spanning tree, there exists one
and only one simple path from v0 to vj . For each vertex in G/V (G) except for
a root v0 , there exists one and only one n such that an0j = 1 and n ≤ m. Depths
of all vertices can be then obtained by computing An while n ≤ m.
⎛ ⎞
a00 a02 a03 a04 a06 a07 a08 a09
⎜a20 a22 a23 a24 a26 a27 a28 a29 ⎟
⎜ ⎟
⎜a30 a32 a33 a34 a36 a37 a38 a39 ⎟
⎜ ⎟
⎜a40 a42 a43 a44 a46 a47 a48 a49 ⎟
A = ⎜
⎜a60
⎟
⎜ a62 a63 a64 a66 a67 a68 a69 ⎟
⎟
⎜a70 a72 a73 a74 a76 a77 a78 a79 ⎟
⎜ ⎟
⎝a80 a82 a83 a84 a86 a87 a88 a89 ⎠
a90 a92 a93 a94 a96 a97 a98 a99
⎛ ⎞
0 1 1 1 0 0 0 0
⎜0 0 0 0 1 1 1 1 ⎟
= ⎜
⎝0 0 0 0 0 0 0 0 ⎟
⎠ (1)
..
.
⎛ ⎞
0 0 0 0 1 1 1 1
A2 = ⎝ 0 0 0 0 0 0 0 0 ⎠ (2)
..
.
A3 = 0 (3)
3 Implementation
3.1 Networking Utilities
In order to control a switch, we have implemented the Networking Utilities that
connect to a switch, and execute a command on a switch. The utilities are written
in Ruby, and can be run with Ruby 1.9.3 or later. The utilities require net/telnet
and net/ssh library, and then connect to a switch by telnet or SSH. The utilities
are free software licensed with 2-clause BSD license, and available on https://
github.com/ohmori7/netutils.
After a network graph is produced and non-reboot switches are detected, vertices
of non-reboot switches are contracted. A depth of each vertex from a root is then
computed, all vertices of switches are grouped by depth. In order to minimize
a delay, all edge switches are logged in before rebooting. The switches of the
deepest vertices are then concurrently rebooted in the deepest vertex first fashion
using threads. The reason why reboots are concurrently not parallely is just
because of a Ruby implementation of threads that can run only one thread at
the same time. When a switch is rebooted, a switch is confirmed to accept a
reboot command if the switch holds one of the following conditions:
After all switches of the deepest vertices are confirmed to start to reboot, the
next deepest vertices are examined.
4 Evaluations
4.1 Evaluation Environment
We have three main campuses, Koyama, Hamasaka, Yonago and tree branch
offices. Koyama and Yonago was connected by 10 Gbps. Koyama and Hamasaka
was connected by 1 Gbps by commercial VLAN service The branch offices were
connected to Koyama by 1 Gbps. The number of network equipment supported
by the Network Utilities in our campus network was 312 in total. 13 switches and
network equipment were then defined as a root. The number of edge switches
to be rebooted was 283, and some switches were not included because they
were shut down during evaluations due to building renovations. All switches
synchronized time using Network Time Protocol (NTP) [2] and its error would
be within 10 ms. All switches also run Rapid Spanning Tree Protocol (RSTP)
per VLAN in order to avoid a loop.
We implemented the Network Utilities on a server running on an ESXi as a
Virtual Machine (VM). We then implemented a code to measure time in Network
Utilities.
158 M. Ohmori et al.
4.4 Downtime
When switches are rebooted, downtime is the most important. We then measured
the downtime defined in Sect. 2.5 by logs of switches. A switch that was firstly
rebooted could log time when a reboot command was received. The time was
regarded as the beginning of downtime On the other hand, the time that a switch
lastly started to forward packets on the last VLAN was regarded as the end of
downtime. This time could be found by RSTP logs. The downtime was then 88 s.
The Deepest Vertex First Reboot 159
Let us compare this downtime and the ideal minimum downtime. As described
in Sect. 2.5, the ideal minimum downtime is to max{vi ∈ G/V (G)|Treboot (vi )}
where Treboot (·) denotes the time that the edge switch, vi , reboots and becomes
up to start to forward packets. In order to find max{Treboot (vi )}, we computed
downtimes of several switches even though they were not all switches. AlaxalA
AX260A-08T, AX2230S-24P, AX2530S-08P, AX2530S-24T, AX2530S-24T4X,
AX2530S-48T and AX2530S-48T2X then took 63, 69, 74, 70, 68, 75 and 73 s to
start to forward packets, respectively. It could be said that max{Treboot (vi )} was
at least more than 75 s. Suppose that 75 s was max{Treboot (vi )}, and it could be
said that the overhead of the deepest vertex first reboot was about 13 s. It can be said
that the deepest vertex first reboot can reboot switches only by approximately 20%
downtime increase. Ones may wonder why 13 s are shorter than 18.61 s as shown
in Table 1. The cause would be that the switch requiring more time to reboot was
rebooted earlier than other switches at the same depth.
Note that investigating the actual max{Treboot (vi )} and more detailed anal-
yses are future work.
5 Concluding Remarks
This paper has proposed the deepest vertex first reboot that can reboot many
network switches with less overhead downtime. This paper has tried to express
a campus network in a graph theory fashion. The deepest vertex first reboot can
then reduce downtime overhead by vertex contraction. The deepest vertex first
reboot then enables to reboot all switches by only 13-s overhead where the under-
estimated ideal minimum downtime is 75 s in an actual campus network where
there are more than 300 network equipment.
160 M. Ohmori et al.
References
1. Hesselink, W.H.: A classification of the nilpotent triangular matrices. Compos.
Math. 55(1), 89–133 (1985)
2. Mills, D.: Network Time Protocol (Version 3) Specification, Implementation and
Analysis. RFC 1305 (Draft Standard). Obsoleted by RFC 5905, March 1992
3. Pemmaraju, S., Skiena, S.: Computational Discrete Mathematics: Combinatorics
and Graph Theory with Mathematica. Cambridge University Press, Cambridge
(2003)
Analyzing and Recognizing Pedestrian
Motion Using 3D Sensor Network
and Machine Learning
1 Introduction
Research on human movement behavior has a long history and a wide range of
applications, especially in recent years, the research that reproduces human motion
behavior such as computer graphics, games, robots and artificial intelligence has made
remarkable achievements. Walking/running is one of the most characteristic chain
movements. Elucidation of the mechanism of walking and running is extensively
involved in training of athletes, and orthopedic surgery fields that specializes in the
treatment of lower limb disorders such as the foot, knee and hip joints. Many patients
with orthopedic diseases are causing obstacles by mechanical stress of actions centered
on walking. On the other hand, running is one of the most basic movements in sports
and there are various running forms. Each player has his/her own form.
At present, most doctors and coaches use their own naked eyes to observe the
walking/running style of patients or athletes, which lacks objectivity and appropri-
ateness. Therefore, we want to construct and develop a mobile pedestrian motion
analysis system by introducing advanced technology such as sensing and artificial
intelligence, which can provide a scientific analysis method based on accurate personal
data.
Advances in computer graphics for modeling human motion and the development
of sensing technology have led to the emergence of motion capture systems. From the
most simplified devices such as Kinect to large professional systems, motion capture
systems are capable of measuring and recording human motion. However, the problem
with conventional camera-based motion capture systems is that they can capture only a
narrow range of motion in front of the sensor and a large studio is required. They
cannot catch natural movements in a wide range of outdoor activities.
On the other hand, it is easier to convert human motion into digital data using
motion sensors such as acceleration sensor or angular velocity sensor. Although the
motion sensor devices can capture and record human motion, it is difficult to visually
recognize human motion behaviors by sensor only, for example, to determine whether
a person is walking or running by view the sensing data only.
Analysis of the motion data from the camera or the depth sensor is based on image
processing technology, while the analysis of the motion data captured by the accel-
eration sensor or angular velocity sensor is based on signal processing. Both have their
own strengths and weaknesses, by blending them we can make use of strengths and
complement the weaknesses. Recently the extraction and classification of sensor data
has also attracted people’s attention, and its importance has been understood year by
year. Apply artificial intelligence such as machine learning and deep learning to handle
large amounts of sensing data for detecting and recognizing human behavior is our
important subject.
In this study, we aim to develop a wearable mobile motion capture system by using a
depth sensor with a small amount of sensing data for modeling, and tracking human
movement using acceleration sensors and angular velocity sensors. We used data
acquired by the sensor network and machine learning to identify and recognize behavior
of pedestrian motion and tried to explore ways to improve the accuracy of recognition.
In this paper, we first describe how to get data from the sensor network and how to
process raw data. Next, we provide algorithms for applying machine learning to the
analysis and recognition of pedestrian motions. Finally, we give our experimental
results and discuss the possibility using our methods to analyze and recognize human
motion.
2 System Overview
We have been building a system in accordance with the following two points:
(1) collecting and storing pedestrian motion data through the sensor network, (2) using
machine learning to analyze motion data and identify pedestrian movements. Here we
give an overview of our system.
When using a Kinect, the motion that can be taken is limited because when the
bone to be tracked enters the blind spot of the sensor, or when the body parts overlap,
the follow-up cannot be successfully followed. In order to overcome the deficiencies of
Kinect’s tracking of stereo motion and improve accuracy, we propose a method of
arranging multiple sensors to complement each other’s blind spots and perform motion
capture as shown in the left of Fig. 1. We combine and complement the obtained
position data so that it can cope with more complex motions [1].
TSND121
TSND121
TSND121
TSND121
TSND121
Fig. 1. (The left) Arranged Kinect pair, (the right) walking experiment with TSND121
Here we explain respectively how to obtain motion data from two kinds of sensors,
Kinect and TSND121, and how to sample the data for analysis.
3.1 Method 1
Kinect can detect 25 skeleton joints and track motions of up to six models at same time.
And each skeleton joint information includes such as jointType, three-dimensional
position coordinate, and tracking status (Tracked, Inferred, NotTracked), which are
updated at 30 fps to track motion. Horizontal angle is 70°, vertical angle is 60°, and
detectable distance is 0.5–4.5 m [1].
In our experiment we set a pair of Kinects to compensate blind spot as the left of
Fig. 1, we called them A and B. When shooting motion with two Kinects, the coor-
dinate set of the motion data obtained by each Kinect is represented on the coordinate
system A and B unique to each Kinect. It is necessary to unify the respective coordinate
systems before completing the coordinate of skeleton joint. According to the coordinate
system of a virtual monocular camera, we a apply series of affine transformations to the
coordinates set of each motion data [5].
Analyzing and Recognizing Pedestrian Motion Using 3D Sensor Network 165
Transformation Algorithm
Step 1. Translate DA to the coordinates of FootLeft of DA1 overlap the origin of the
virtual monocular camera.
Step 2. Find out the angle a about the y axis that the z component of the position
vector of FootRight becomes 0 and rotate DA around the y axis a degree.
Step 3. Find out the angle b about the z axis that the y component of the FootRight’s
position vector becomes 0 and rotate DA around the z axis b degree.
Step 4. Find out the angle c about the x axis that the z component of the position
vector of the Head becomes 0 and rotate DA around the x axis c degree.
Step 5. The final coordinates set D0A is obtained by (1), where v is the position vector
of the FootLeft of DA1 and R is the rotation matrix (2),
As the result of the above process, the FootLeft is on the origin, the FootRight is on
the x axis, and the Head is in contact with the xy plane.
We give an example in Fig. 3, where the red model captured by Kinect A and the
blue model captured by Kinect B. In the images of A and B, the yellow skeleton joint
indicates that its tracking state is “inferred” that means the joint cannot be captured
correctly. Some “inferred” joints appear in red but not in blue, and vice versa. We
166 N. Sun et al.
proposed an algorithm that complements these hidden joints [1]. The complemented
result in the most right of Fig. 3 is used for our tracking and motion analysis.
3.2 Method 2
In each action experiment with motion sensor TSND121, we sampled three times, three
seconds each respectively, collected a total of 9000 data. Due to the efficiency of
Bluetooth communication between the server and the motion sensor, the amount of data
acquired was less than or greater than 9000 sometimes. In our experiments we sampled
data of walking and running, and each responding six-axis information included
acceleration (Ax , Ay ; Az ) and angular velocity (xx , xy ; xz ) acquired from the waist, left
thigh and left lower leg, right thigh and right lower leg like the right of Fig. 1.
The angle can be calculated by integrating the angular velocity with time. Usually
the measured value of the sensor contains an error. If we integrate as they are, the error
will accumulate, and the correct result will not be obtained. On the other hand, when
the angle is continuously obtained over a long period of time using the acceleration,
there is almost no error, but the error is large when they are obtained in a short time.
When angular angles are used to obtain angles continuously for a long period of time,
there are many errors, but the errors are small when calculated in a short time [6].
Therefore, we only extract the good part of both acceleration and angular velocity and
find the angles by complementing each other.
Let u be the roll angle of the sensor, h be the pitch angle, and t be the time.
Calculate ua and ha from the acceleration ðAx , Ay ; Az Þ by the (3) and (4).
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ua ¼ atan2 Ay ; A2x þ A2z ð3Þ
ha ¼ atan2ðAz ; Ax Þ ð4Þ
The motion angles are calculated using the complementary filter (5) and (6) where
R
xx dt is roll angle and R xy dt is the pitch angle obtained from angular velocity, ui is the
corrected roll angle and hi is the corrected pitch angle, where k = 0.95,
1 i n. Figure 4 gives the corrected results of walking experiment.
Analyzing and Recognizing Pedestrian Motion Using 3D Sensor Network 167
Z
ui ¼ k ui1 þ xx dt þ ð1 k Þua ð5Þ
Z
hi ¼ k hi1 þ xy dt þ ð1 kÞha ð6Þ
Fig. 4. Corrected roll (the left) and pitch (the right) angles of walking
For the above Methods 1 and 2, we proposed some methods for analyzing and rec-
ognizing the motion behavior of the measured object using sensing data and machine
learning. In this section, we present our algorithms and approaches respectively.
4.1 Method 1
Recognition Using Posture Similarity. We chronologically vectorize the two pose
coordinate sets, one set of a and one set of b, and apply (7) to find the Euclidean
distance d between the joints of a and b. From our experiment results, we defined a
small threshold value e = 8. If the Euclidean distance d of the two poses is less than or
equal to the threshold e, the two poses are judged to be the same posture, otherwise it
will be judged as a different posture. If there are two or more tests in which the
Euclidean distance d for different motion is less than or equal to e, they are treated as
unrecognizable. Learning and recognition can be achieved by simply comparing the
test data set b with a previously prepared teacher pose model a. Accuracy can be
improved by preparing approximately 50 learning data and make the average of each
element as the model of the posture a.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n
d¼ ðai bi Þ2 ð7Þ
i¼1
that the nature of the posture can be represented by images, and the image conversion
operation does not affect the learning effect. We used Sony Neural Network Console
(SNNC) for construction, learning and evaluation of the convolution neural network.
First, it is necessary to convert the vector of the posture into a color image data. We
translated and scaled values of all the elements of the vector, so that values fall within
the range of 0 to 255, then the coordinate (x, y, z) are converted into one pixel as the (r,
g, b) pixel values. The learning process consists of five steps: (1) Input data of learning
and evaluation to the neural network. In this study, 50 images were prepared for each
posture as training data for learning, and the other 25 data were input for evaluation.
(2) Convolution: the size of the filter is 6 3, and the initial weight value of the filter
pffiffiffiffiffiffiffiffi
is w ¼ randð 2=nÞ devised by Kaiming He. This initial value was devised as an
optimum value when ReLU to be described later is used. (3) ReLU: add ReLU (ramp
x; x 0
function) as network activation function that is Rð xÞ :¼ . (4) Affine: the input
0; x\0
data is converted into a sequence of 100 elements by affine transformation. (5) Soft-
maxCrossEntropy: the data is classified using the Softmax function and cross entropy,
and the results are output.
4.2 Method 2
Feature Values Definition. Generally, pedestrians walk or run forward, so we only
use the acceleration Az from the acceleration (Ax ; Ay ; Az ) for analysis shown in red
curves of Fig. 5. In addition, the angular velocity xy plays an important role in all
pedestrian movement. In walking, stepping, and running, angular velocity xy repre-
sents the upward and downward angular velocity in the vertical direction shown in blue
curves of Fig. 5. We defined the standard deviation, maximum, and minimum of the
acceleration Az and angular velocity xy as the feature value in our analysis, which is
abbreviated as SDMM that is a 6-dimensional vector [3].
Feature Values Calculation. The peak waveform of Az and xy in Fig. 5 tell us that a
person can complete a step in about 700 ms. Therefore, we separate the sampling
interval by 700 ms. We further divided each part into two and made two windows of 0
to 350 ms and 350 to 700 ms. In doing so, the portion of the right thigh that is raised in
one step and the portion where the leg is lowered can be separated as a separate
window. We compute the SDMM for each of the two split windows and combine the
two partial SDMMs into the final feature SDMM of the 12-dimensional vector, which is
a row of the training data set or test data set. Next, we move the window from the
previous window 50 ms to the next 700 ms in order. Again, we calculate SDMM for
this new window. An integer n indicates the total number of rows in the training data
set or test data set in (8). In our experiments, when the training data was acquired for
3 s, there will be approximately 46 rows of SDMM for 3000 sets of acceleration z and
angular velocity y, with 12 columns per row, so up to 552 vectors should be calculated.
Fig. 5. Samples of walking (the left) and running (the right) motion data
and 2 give the recognition results. F-measure of similarity is 0.88, SNNC is 1, which
tell us that converting the coordinate (x, y, z) of the postures into the color (r, g, b) of
the images make it becomes possible to use good efficient SNNC to recognize the
motion.
Evaluation Results of Method 2. We verified the method of 4.2 using three types of
movement: step, walk, and run. For each movement we prepared 9000 training data,
and one set of test data in which the first half is the data obtained from walk and the
second half from run respectively. We want to find out how close the test action type is
to already the given training types with (9) and k-NN classifier.
Figure 6 shows that recognition method of 4.2 has better recognition rate. Com-
pared to Method 1, Method 2 is simplified and easy-understanding. However, for
different measurement objects, the size of the sizing window used to divide the
waveform data samples affects the recognition accuracy.
As a solution, trial and error is performed to increase the learning data and adjust
the sizing window size of the raw data and prepare appropriate learning features set
before recognition.
Analyzing and Recognizing Pedestrian Motion Using 3D Sensor Network 171
Fig. 6. Recognition results of walking and running with (9) and k-NN
5 Conclusion
References
1. Sun, N., Murakami, S.: Human motion modeling from complementary skeleton joints of
multiple kinects. In: Proceedings of the 2018 International Conference on Mathematics,
Modelling, Simulation and Algorithms. Advances in Intelligent Systems Research, vol. 159,
pp. 131–135 (2018)
2. Sun, N., Sakai, Y.: New approaches to human gait simulation using motion sensors. In: 31st
International Conference on Advanced Information Networking and Applications Workshops
(WAINA) (2017). 16901760
3. Sun, N., Tsuruoka, T.: Pedestrian action recognition using motion sensor and k-nn classifier.
In: Proceedings of 2nd International Conference on Artificial Intelligence: Technologies and
Applications. Advances in Intelligent Systems Research, vol. 146, pp. 1–4 (2018)
4. Miyajima, S., Tanaka, T., Miyata, N., Tada, M., Mochimaru, M., Izumi, H.: Feature selection
for work recognition and working motion measurement. J. Robot. Mechatron. 30(5), 706–716
(2018)
5. Foley, J.D., Cardelli, L., van Dam, A., Feiner, S.K., Hughess, J.F.: Computer Graphics:
Principles and Practice, 3rd edn., pp. 263–286, 299–320. Addison-Wesley, New York (2015)
6. Lambrecht, S., Nogueira, S.L., Bortole, M., Siqueira, A.A.G., Terra, M.H., Rocon, E., Pons, J.
L.: Inertial sensor error reduction through calibration and sensor fusion. Sensors 16(2), 1–16
(2016)
Novel Interestingness Measures
for Mining Significant Association Rules
from Imbalanced Data
Safa Abdellatif1(B) , Mohamed Ali Ben Hassine1 , and Sadok Ben Yahia1,2
1
Faculty of Sciences of Tunis, University of Tunis El Manar,
LIPAH-LR11ES14, El Manar, 2092 Tunis, Tunisia
{Safa.abdellatif,mohamedali.benhassine}@fst.utm.tn
2
Department of Software Science, Tallinn University of Technology,
Akadeemia tee 15a, 12618 Tallinn, Estonia
sadok.ben@taltech.ee
1 Introduction
In the last decade, with the huge advance of technology and fast communication,
enterprises are daily collecting and storing large amounts of data from various
sources. In order to handle such overwhelming data by discovering unknown and
potentially useful information, several data mining techniques have been shown
to be of benefit. One of the most popular data mining techniques is Association
Rule Mining (ARM). It aims to detect relevant and valuable associations between
items in a database.
Associative classification (AC) is the integration of ARM and classification
for the purpose of constructing classifiers to predict the class labels of the new
data. Roughly speaking, ACs use predictive association rules which have as a
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 172–182, 2019.
https://doi.org/10.1007/978-3-030-15035-8_16
Novel Interestingness Measures for Mining Significant Association Rules 173
Lenca et al. [5] have proposed five properties for the IM evaluation: (i) L1 . IM
is constant if there is no counterexample to the rule. L1 states that if P (AC) = 0
then IM’s value is constant. In other words, the rule with a value of confidence
equal to one should have the same interestingness value regardless the support
which contracts another property of Major and Mangano [7] that states that a
measure should combine support and association aspects; (ii ) L2 . IM decreases
with P (AC) in a linear, concave or convex fashion around 0+ . As a few numbers
of counterexamples are added, the L2 property describes in which manner the
IM decreases.
L3 . IM increases as the total number of records increases. L3 states that the
value of IM increases with the total number of records; (iii) L4 . The threshold
is easy to set. L4 states that it is easy to choose a threshold that separates
interesting from uninteresting rules; (iv ) L5 . The semantics of the IM is easy
to express: This property states that it is desirable that the semantics of the
Novel Interestingness Measures for Mining Significant Association Rules 175
On the other hand, we can also prove that the lift measure favours rules
from minor classes. An association rule is considered as important if the lift value
highly surpasses 1. For major classes C1 , the lift value is defined as lif t(A → C1 )
sup(AC1 )
= sup(A)sup(C 1)
. Since sup(C1 ) 1 and sup(AC1 ) sup(A) then lif t(A →
C1 ) 1.
As a result, the lift measure may hardly find any interesting rule from major
classes but it is beneficial for the extraction of rules from minor classes. However,
one of the major drawbacks of the lift measure is its symmetry. In fact, the lift
measure takes A and C in equivalent position i.e. lif t(A → C) and lif t(C → A)
are the same, which is not always true in case of imbalanced datasets.
We propose several measures which aim to capture rules which are beneficial for
major and minor classes. The first proposed IM is called M odif iedLif t and is
defined as follows:
lif t(A → C) P (A) P (AC)
M odif iedLif t(A → C) = ∗ = (1)
lif t(A → C) P (A) P (AC)
The first part of the formula refers to the ratio of the lif t measure of A → C
divided by the lif t measure of A → C. As the value of lif t(A → C) increases, A
and C are dependent. Consequently, there is a strong possibility that A and C are
dependent too. Moreover, the higher the lif t(A → C) is, the worse the rule A →
C. For these reasons, we have proposed lif t(A → C) as nominator and lif t(A →
C) as denominator. However, this ratio tends to have very high values. So we
propose to make some corrections by multiplying this ratio by the ratio of the
no occurrence possibility of premise divided by the occurrence probability of the
premise. As a conclusion, we may notice that the M odif iedLif t is getting larger
when (no occurrence of A implies no occurrence of C) surpasses (occurrence of
A implies the occurrence of C). This measure could be beneficial in case of
imbalanced datasets. In fact, if we take the example of a binary imbalanced
dataset with C1 as the major class and C2 as the minor one. The value of
Novel Interestingness Measures for Mining Significant Association Rules 177
since for the minor class C2 , we have P (AC2 ) = P (C2 ) and P (AC2 ) = P (C2 ).
On the other hand, the value of M odif iedLif t for the rule A → C2 is equal to:
Based on these approximations, we may notice that the value of M odif iedLif t
for rules extracted from major classes is around one. However, for rules extracted
minor classes, this value depends on both the occurrence and non occurrence of
the premise part. We may notice that for this type of classes and based on our
approximations, we have:
1 P (AC)
DM1 (A → C) = M odif iedLif t(A → C) ∗ = (2)
P (AC) P (AC)P (AC)
1 P (AC)
DM2 (A → C) = M odif iedLif t(A → C) ∗ = (3)
P (C) P (AC) P (C)
could be suitable for imbalanced datasets since it favors rules having low P (A)
which generally correspond to long antecedents. DM4 is defined as follows:
1 P (AC)
DM4 (A → C) = M odif iedLif t(A → C) ∗ = (5)
P (A) P (AC) P (A)
Measure Properties
P1 P2 P3 L1 L2 L3
ModifiedLift No Yes Yes No Yes Yes
DM1 No No Yes No Yes Yes
DM2 No Yes Yes No Yes Yes
DM3 No Yes Yes No Yes Yes
DM4 No Yes Yes No Yes Yes
Confidence No Yes Yes Yes Yes No
Laplace No Yes Yes No Yes Yes
Lift Yes Yes Yes No Yes No
Conviction Yes Yes Yes Yes Yes No
4 Experimental Results
4.1 Data Collection and Experimental Setup
In order to evaluate the efficiency of the newly proposed IMs, we have selected
six real-world datasets: five of them are taken from the UCI repository [8] and
the last one is composed of posts collected from Twitter. The characteristics
of these datasets are summarized in Table 2. For each dataset, the number of
instances, the number of attributes and the imbalance ratio are provided. The
latter is defined as the ratio of the number of instances of the major class divided
by the number of instances of the minor class.
five new measures will be compared to the adapted ARCID with four existing
measures (i.e. Confidence “Conf”, Laplace “Lap”, Lift and Conviction “Conv”)
and several well-known classification approaches including CBA [6], C4.5 [10]
and Ripper [3]. Experimentation will be carried with reference to four assessment
measures which are the Global Accuracy, Geometric Mean, Weighted Accuracy
and F-measure.
F-measure. Table 6 presents the f-measure results obtained from the experi-
ments. It shows that the ARCID versions using the newly proposed measures
outperform all the rest terms of F-measure which owe to the fact that it asserts
an absence of a bias towards minor classes.
Considering all of the above, we may conclude that the ARCID approach
using the newly proposed measures gets better ranks than ARCID using the
existing measures as well as the state-of-art algorithms and this is according
to the four evaluation measures used. To sum up, we may conclude that the
newly proposed measures are beneficial in case of imbalanced datasets since they
efficiently extract significant knowledge from minor classes (proven using the
Geometric mean, the Weighted Accuracy and the F-measure) without decreasing
the predictive accuracy of the classifier (proven using Global Accuracy metric).
Novel Interestingness Measures for Mining Significant Association Rules 181
5 Conclusion
This paper deals with the problem of association rule selection for prediction
in case of imbalanced datasets. Existing measures do not yield good results in
case of imbalanced datasets since they tend either to favour rules of major class
and consider others as noise or emphasize only on the rules of minor classes and
omit others. This paper proposes five new IMs to keep highly interesting rules
from both types of classes simultaneously in order to be able to offer a high
recall on the minority class while maintaining a high precision of the majority
class. Experiments show that the newly proposed measures offer a better pre-
dictive power compared to the existing ones. In our future work, we intend to
study the predictive power of the IMs with several datasets in text mining and
sentiment analysis. Moreover, we plan to apply these new IMs on well-known
algorithms such as CBA [6] in order to make them suitable for the classification
of imbalanced datasets.
182 S. Abdellatif et al.
References
1. Abdellatif, S., Ben Hassine, M.A., Ben Yahia, S., Bouzeghoub, A.: ARCID: a new
approach to deal with imbalanced datasets classification. In: International Confer-
ence on Current Trends in Theory and Practice of Informatics. Springer (2018)
2. Abdellatif, S., Ben Yahia, S., Ben Hassine, M.A., Bouzeghoub, A.: Fuzzy aggre-
gation for rule selection in imbalanced datasets classification using choquet inte-
gral. In: 2018 IEEE International Conference on Fuzzy Systems, FUZZ-IEEE 2018,
Rio de Janeiro, Brazil, 8–13 July 2018 (2018)
3. Cohen, W.W.: Fast effective rule induction. In: Proceedings of the Twelfth Inter-
national Conference on Machine Learning, pp. 115–123 (1995)
4. Hu, B.G., Dong, W.M.: A study on cost behaviors of binary classification measures
in class-imbalanced problems. arXiv preprint arXiv:1403.7100 (2014)
5. Lenca, P., Vaillant, B., Meyer, P., Lallich, S.: Association rule interestingness mea-
sures: experimental and theoretical studies. In: Quality Measures in Data Mining,
pp. 51–76. Springer (2007)
6. Ma, Y., Hsu, W., Liu, B.: Integrating classification and association rule mining.
In: Proceedings of the Fourth International Conference on Knowledge Discovery
and Data Mining (1998)
7. Major, J.A., Mangano, J.J.: Selecting among rules induced from a hurricane
database. J. Intell. Inf. Syst. 4(1), 39–52 (1995)
8. Merz, C.: UCI repository of machine learning databases (1996). http://www.ics.
uci.edu/∼mlearn/MLRepository.html
9. Piatetsky-Shapiro, G.: Discovery, analysis, and presentation of strong rules. In:
Knowledge Discovery in Databases, pp. 229–238 (1991)
10. Quinlan, J.R.: C4.5: Programs for Machine Learning. Morgan Kaufmann,
San Francisco (1993)
Extended Scheme Mediation Integration Model
for Information Systems Project Proposal
1 Introduction
Information Systems (IS) face two major challenges in their operation. On an hand,
there is a need to ensure interoperability between the heterogeneous systems that are
developed, in general, independently based on a single domain representation. On the
other hand, the operational environment is dynamic which leads to continuous changes
in these systems. Thus, it is necessary to optimize the process of reusing the modules
between the different systems to accelerate the developments. In analyzing these
challenges, we find that the problem of semantic sharing is a major factor that degrades
interoperability and prevents reuse between ISs.
The possible way to deal with the problem of semantic heterogeneity is to reduce or
even eliminate terminological and conceptual incompatibilities. Therefore, establishing
a common understanding (terminological and conceptual), with multiple points of
view, can help establish a communication base between users, manage interoperability
between systems and improve the engineering process of reuse.
Over the centuries, the constant evolution of communication technologies has made
the sharing human impetus increasingly more important. Sharing information is not a
new idea and in this paper, we focus on some challenging interoperability issues that
emerged with the evolution of communication technologies.
Neither the sharing of information between people nor the exchange of data
between information systems are new ideas. Nevertheless, new challenges are created
by the improvement of interconnection technologies of computational agents and
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 183–192, 2019.
https://doi.org/10.1007/978-3-030-15035-8_17
184 W. C. de Souza Carvalho et al.
improved algorithms for processing and retrieval of large volumes of data. The solu-
tions must meet different requirements and criteria to accomplish the user needs [1]:
autonomy of the systems, extensibility of the architecture to control the addition or
removal of IS, scalability of the system capacity faced with the increase on the users
amount and transparency regarding location and data format. The solutions are based
on the notion of interoperability, that is, the implementation of a such collaboration
between systems that allows data and service sharing to respond to information
requests. Resolving a query in an interoperable environment involves integrating dis-
tributed data across multiple heterogeneous systems raises some questions:
1. Which ISs may contain all or part of the desired information?
2. What data are relevant to the expressed needs?
3. How to transform, to adapt and to pair the data in order to obtain a consistent and
user-friendly result?
The information integration is preceded by resolution of three types of conflicts:
• Syntactic conflicts result from using different data models from one system to
another; beyond what, different concepts are used to structure the same information.
• Schematic conflicts arise from different structures and classifications of information.
This is closely related to the choices made during designing the project.
• Semantic conflicts arise from differences in the interpretation of information shared
by various application domains: nomenclature conflicts (taxonomic and linguistic
problems), value conflicts (unit problems, scales), and cognitive conflicts (meaning
problems).
Schematic and semantic conflicts are closely related. The meaning of information
must be clearly defined, and semantic conflicts resolved before schematic conflicts are
identified and addressed.
Many solutions have been proposed to consider these data conflicts, that guarantee
at least the autonomy, the composability and the resolution of syntactic conflicts.
Multi-base approaches rely on the use of a query language and a common repre-
sentation model [2]. Each SI exports its information as a schema described in the
common model (usually an object-oriented model), the multi-base language, extended
from SQL (Structure Query Language) or OQL (Ontology Query Language), allows a
multi-site query. The multi-base approach is extensible but does not offer localization
transparency. The resolution of schematic and semantic conflicts remains entirely on
user’s responsibility.
Federated approaches rely on integration. Each SI exports a schema in a common
model, so the various exported schemas are integrated into federated schemas. A fed-
erated schema allows access to shared data in a uniform and global way [3]. Schematic
and semantic conflicts are solved by the integration process. Transparency is ensured,
just by posing a query on the federated schema. However, extensibility and scalability
are criteria that are not adequately met; which indicates that federation is an approach to
be considered when integration involves a small number of information systems and
when these systems go through few evolutions.
Mediation approaches extend the federation by providing more flexibility.
Mediation rely on an essential component, named mediator, which is responsible for
Extended Scheme Mediation Integration Model 185
meeting the integration needs based on the knowledge made available [4, 5]. The
mediator finds the available information and solve the detected schematic and semantic
conflicts. A secondary component serves as interface to the ISs and solve syntactical
conflicts by providing data to the mediation model.
Two types of mediation can be distinguished according to how data conflicts are
resolved:
1. Mediation of schemes that previously builds a database of information about the
participating ISs which gives the mediator the means to do his integration work.
2. Context mediation in which the mediator uses semantic information to resolve
queries dynamically without prior knowledge of which information systems belong
to the context.
Schema mediation is a direct extension of the federated approach [1, 6, 7] with better
scalability and often better scalability (object interfaces, rules-based language) [8, 9].
Context mediation seeks to discover data semantically close. This such mediation can
find and adapt information to ensure total transparency [10–13, 21, 22, 27, 28].
This paper presents the main aspects of a under development extended scheme
mediation integration model for information systems project proposal. When finalized,
the model will be composed by a methodology, a model and an architecture to build a
knowledge base to integrate information corresponding to well-defined usage profiles.
The mediation model will make it possible to represent the data structure in a common
model, adapt and compose the information in order to ensure total transparency. This
model will also associate semantics with shared information to compare and identify
relevant data. An incremental integration mechanism guided by semantics will allow its
adaptation to dynamic and broad environments.
Section 2 provides an overview of solutions based on context mediation and then
introduces the key points of for a integration model based on extended scheme medi-
ation. Section 3 describes a typical example of a cooperation environment and explains
how our approach can solve problems related to interoperability of IS. Section 4 con-
cludes the article and presents research perspectives.
2 Related Work
3 Semantic Approach
all the knowledge allowing to give a meaning to the concept, the taxonomic infor-
mation (synonyms, antonyms), the properties (semantics of values, constraints) and the
relations with the neighborhood. The signified is the conceptual class that represents
the abstraction of an entity.
As SIs can exports mediation model object classes, which provide a local context,
these classes could be defined in relation to their local context, so the mediation model
will describe both the structure and the semantics of the data. In other hand, semantic
reconciliation between the contexts of two SIs allows to import the relevant classes and
integrate them.
The information offered by this association includes for each concert the name of
the artist, the date of the concert, the number of places still free, the price of places and
the room where it takes place.
A user can query this system to obtain an integrated view of the information
distributed on various information sources. To simplify the example, we consider the
existence of two systems s1 and s2 relating to two concert halls and likely to give
information on concerts.
Extended Scheme Mediation Integration Model 189
In this paper, we have presented an extended scheme mediation integration model for
information systems project proposal. Our approach uses a model that describes the
shared information by taking their semantics into account in the context of usage
contexts. The definition of local contexts location and use of shared and heterogeneous
information turn our work in the direction of further formalizing the notion of semantic
distance and moving towards the implementation of architecture. The next steps of
creating our model will evolve tailoring a methodology based on contextual matching
and flexible architectures which will allow incremental integration of relevant infor-
mation in the context of information systems. Further steps will also be responsible to
formalize the definition of local contexts, contextual matching and semantic distance to
realize the interpretation, finding and use of shared and heterogeneous information. Our
current work goes in the direction of further formalizing the notion of semantic distance
and moving towards the implementation of an architecture.
Extended Scheme Mediation Integration Model 191
References
1. Liu, L., Pu, C.: Approach to large-scale interoperable database systems. Technical report
TR95-16, University of Alberta (1995)
2. Floresco, D., Rachid, L., Valdurez, P.: Using heterogeneous equivalences for query rewriting
in multidatabase systems. In: Proceedings of the Third International Conference on
Cooperative Information Systems (1995)
3. Sheth, A.P., Larson, J.A.: Federated database systems for managing distributed, heteroge-
neous, and autonomous databases. ACM Comput. Surv. 22(3), 183 (1990)
4. Wiedershold, G.: Mediators in the architecture of future information systems. IEEE Comput.
Mag. 25(3), 38 (1992)
5. Wiedershold, G.: The basis for mediation. In: Proceedings of the Third International
Conference on Cooperative Information Systems (1995)
6. Cody, W.F., Has, L.M., Nblack, W., Arya, M., Carey, M.J., Fagim, R., Flickner, M., Lee,
D., Petkovick, D., Schwarz, P.M., Thomas, J., Roth, M.T., William, J.H., Wimers, E.L.:
Querying multimedia data from multiple repositories by content: the garlic project. In:
Proceedings of the Third IFIP (1995)
7. Tonasic, A., Rasschid, L., Valdurez, P.: Scaling heterogeneous databases and the design of
disco. In: Proceedings of the IEEE 16th International Conference on Distributed Computing
Systems. IEEE Computer Society Press (1996)
8. Garcia Molina, H., Quas, D., Papakostantinou, Y., Rajarman, A., Sagiv, Y., Ullman, J.D.,
Widon, J.: The TSIMMIS approach to mediation: data models and languages. In:
Proceedings of the Second International Workshop on Next Generation Information
Technologies and Systems (1995)
9. Papakonstatinou, Y., Abitebul, S., Garcia Molina, H.: Object fusion in mediator systems. In:
Proceedings of 24th International Conference on Very Large Data Bases (1996)
10. Kashyap, V., Sheth, A.: Semantics-based information brokering. In: Proceedings of the 3rd
International Conference on Information and Knowledge Management, Gaithersburg. ACM
Press (1994)
11. Bishr, Y.: Semantic aspects interoperable GIS. Ph.D. thesis, Enschede Netherlands (1997)
12. Bressam, S., Goh, C.H., Fiynn, K., Jakobisiack, M., Husein, K., Kon, H.B., Lee, T.,
Maddnick, S.E., Pena, T., Qu, J., Shum, A.W., Siegel, M.: Context interchange mediator
prototype. In: Proceedings of ACM SIGMOD International Conference on Management of
Data (1997)
13. Deker, S., Erdmamn, M., Fensel, D., Studer, R.: Ontobroker: ontology based access to
distributed and semi-structured information. In: Eighth Working Conference on Database
Semantics, DS-8, Kluwer (1999)
14. Grubber, T.: A translation approach to portable ontology specifications. Int. J. Knowl.
Acquisition Knowl. Based Syst. 5(2), 199–220 (1993)
15. Guarino, N.: Semantic matching: formal ontological distinctions for information organiza-
tion, extraction, and integration, summer school on information extraction, Frascati, Italy,
July 1997
16. Kiffer, M., Lausem, G., Wu, J.: Logical foundations of object-oriented and frame-based
languages. J. ACM 42(4), 741 (1995)
17. Mena, E.: OBSERVER: an approach for query processing in global information systems
based on interoperation across pre-existing ontologies. Ph.D. thesis, University of Zaragoza
(1999)
18. Fowler, J., Perry, B., Bargmeyer, B.: Agent-based semantic interoperability in InfoSleuth.
ACM SIGMOD Rec. 28(1), 60 (1999)
192 W. C. de Souza Carvalho et al.
19. Nicolle, C., Cullot, N., Yetongnon, K.: A translation process between cooperating
heterogeneous database systems. In: Proceedings of PDCS97, New Orleans (1997)
20. Grumbach, A.: Cognition artificielle. Du réflexe à la réflexion. Addison-Wesley France
(1994)
21. Vitvar, T., Kopecky, J., Viskova, J., Fensel, D.: WSMO-Lite annotations for web services.
In: Bechhofer, S., Hauswirth, M., Hoffmann, J., Koubarakis, M. (eds.) ESWC 2008. LNCS,
vol. 5021, pp. 674–689. Springer, Heidelberg (2008)
22. Farrell, J., Lausen, H. (eds.): Semantic annotations for WSDL and XML schema. W3C
Recommendation (2007)
23. McGuinness, D.L., van Harmelen, F.: OWL web ontology language overview. W3C
Recommendation (2004)
24. Hibner, A., Zielinski, K.: Semantic-based dynamic service composition and adaptation.
IEEE SCW 2007, 213–220 (2007)
25. Gomadam, K., Verma, K., Brewer, D., Sheth, A.P., Miller, J.A.: A tool for semantic
annotation of web services. In: Gil, Y., Motta, E., Benjamins, V.R., Musen, M.A. (eds.)
ISWC 2005, vol. 3729. Springer, Heidelberg (2005)
26. Gmez-Prez, J.M., Erdmann, M., Greaves, M., Corcho, O., Benjamins, R.: A framework and
computer system for knowledge-level acquisition, representation, and reasoning with process
knowledge. Int. J. Hum. Comput. Stud. 68, 641 (2010)
27. Krotzsch, M., Simancik, F., Horrocks, I.: Description logics. IEEE Intell. Syst. 29, 12–29
(2014)
28. Kurniawan, K., Ashari, A.: Service orchestration using enterprise service bus for real-time
government executive dashboard system. In: Proceedings of 2015 International Conference
on Data and Software Engineering, ICODSE 2015 (2016)
29. Parlanti, D., Pettenati, M.C., Bussotti, P., Giuli, D.: Improving information systems
interoperability and flexibility using a semantic integration approach. In: Proceedings - 4th
International Conference on Automated Solutions for Cross Media Content and Multi-
Channel Distribution, Axmedis 2008 (2008)
The 15th International Workshop on
Heterogeneous Wireless Networks
(HWISE-2019)
An Analysis on Recent Mobile Application
Trend in Bangladesh
Abstract. Development in the technology sector, making the world closer and
easier day by day. It is very difficult to stay aside without the touch of the
technology. From the invention of computer, the daily life of smart people has
changed. Now we are seeing smart phone is ruling the world. It has replaced the
great invention of computer. Almost every work of a computer and many fea-
tures are now enabled in a small smart phone. The apps inside a smart phone are
one of the main factors that is helping the people to make their life simpler.
Because of the immense usage of smart phone, app developers are increasing
day by day. In this paper, we have tried to find out the recent app using trend in
Bangladesh. We have operated a survey. From there we have figured out the
most demanding app types and also the specification in Bangladesh. Our work
will help the app developers beyond any platform. This will help them to build a
demanding app and be in a good position in the market place.
1 Introduction
The number of mobile phone users in twenty years back and today are so much
amusing that it is pretty much difficult to say how a small device got so much popu-
larity among the all classes of people [1]. In past, it was so costly for a person to buy a
phone. But with time everything has changed. As the cost has went down and the
reachability to phones upgraded, people are willing to buy a phone. The features in a
mobile phone is updating day by day and the technology got a revolutionary change
when it got the taste of first smartphone. How features and technology can be improved
to make the peoples life easy, a smartphone is the best example.
People are getting dependent on tech world to make their life easier and enter-
tainable. In recent times, people want a fast and reliable device for their communication
which would be easy to use as well as the best thing to pass their leisure time. So,
people are changing their choices for communication to smartphones from analog
cellular phones. The main reason of this change is for some eye-catching applications
in the smartphones which are not supported in analog phones. And day by day the
applications are getting more enjoyable and more reliable with the increasing rate of
development of smartphones.
The technology will develop every day and the mobile applications should also be
developed to cope up with the recent running race. In 2014, the number of smartphone
users was 1.57 billion and in 2017, it is 2.32 billion and it is expected that the users will
be 2.87 billion by 2020 [2]. There are many factors behind this increasing number of
users year by year. One of this factor is the mobile applications. People are buying
smartphones to use these applications. So, these applications can be improved or
similar kind of apps can be made.
There are some key factors which are responsible for making a product famous or
popular in the marketplace [3]. The main factors are:
• Knowledge about the targeted product
• Knowledge about the target users’ choice
• Knowledge about the targeted marketplace
Every developer wants to make a good marginable profit from his or her product. It
will be risky to make a product which is not wanted by the user or by a specific user
group. It may cause loss of his or her valuable effort. If a mobile app developer is
willing to build an app then it would be a great loss for the latest up growing
technology.
A mobile application developer needs some knowledge before building a new
app. The first fact to make a new application is the idea because if the app doesn’t get
popular among the users, the effort, time and money everything will be wasted. Maybe
the app is good but the marketplace is not ready for this kind of apps or there can be
another reason that people are wanting something else or there are other similar apps
which are better than this one. And if the developer fails to make his or her app popular,
he or she will lose confidence to work again on this mobile app developing sector.
Our effort is to make the developer’s and freelancers’ life easier. We are focusing to
find out the concurrent technology trend so that a developer can focus only on his or
her app developing work. It will save developer’s valuable time and productivity. By
analyzing the market, we have suggested some most top using apps and their con-
figurations. The developers will have an idea about the app market. They can judge
their idea whether the app they are planning to make will make an impact among the
smartphone users or not. In short, in our work we have achieved:
• Ranked the top usage apps in Bangladesh
• Figured out the category of the top running app
• also figured out which mobile platform is making an impact in profit comparison
• explored out the smartphone configuration which is mostly used by a large number
of people in Bangladesh
• the mobile building giants will also have an idea which configured mobile should be
in the market in near future
The remaining part of this paper is decorated as:
An Analysis on Recent Mobile Application Trend in Bangladesh 197
2 Related Work
The trend of using apps is changing day by day. A certain app is famous for some
certain time. So, the market is fluctuating as the recent technology is jumping to a new
era day by day. So, it is very important to have a good idea about the market to launch a
new product.
There is a research work done on automotive application usage using fuzzy logic
[4]. The focus was mainly on how often the cruise control in cars are used. Summa-
rizing the user’s data into a Cumulative Distribution Function (CDF) and then divided
the speed into four levels, a solution was made using Probability Distribution Function
(PDF) to measure speed preference. And with speed and data the research shows a
result of prediction how often a car owner uses the cruise control in a car.
In a recent research, the categories of mobile applications have been studied [5]. In
that work, four categories are mentioned for various mobile apps. These are - a)
notifications, b) location tracking, c) navigation, d) real-time mobile job scheduling.
Similarly, Habib, Ateeq and Bajwa conducted a survey for examining the basic
trends of the usage of mobile applications and different methods for their allocation on
the mobile phones. To get their desired result, they asked four simple questions through
their survey. The outcome of the survey was successful to measure the popularity of
various internet applications and the distribution methods used for the mobile appli-
cations [6].
Moreover, in another research paper, the authors made an app named ‘Analysis and
Prediction of Application Usage’ (APAU) where their main target was to make
application switching in an android phone smooth, fast and efficient by helping the user
conserve time and effort, which could otherwise be wasted on work other than the
actual usage [7]. They resolved to build a recommendation system that will point out
applications based on the user’s usage pattern. Users can be helped by APAU because
it can accelerate application usage by providing shortcuts to applications that the user
would want to use next.
In the generation we are living, there are tremendous number of apps we can choose
from. People from different sectors can easily access the apps according to their need.
In a work, the researchers had conducted a survey to get the users requirements of an
online course management system [15]. Feng, Worrachananun and Lai proposed a
research based on the people who use apps for their learning purpose. Their study has
analyzed learner’s learning styles and better understands their perceptions on the use of
a smartphone application based on the unified theory of acceptance and use of tech-
nology [8]. Interestingly, they discovered that most of the learners prefer visual
learning style, diagrams and graphs which can help them to gather information and
memorize things.
Similarly, in another research, Diliberto-Macaluso and Hughes brought in a
resembling idea with a little modification. The goal of their experiment was to deter-
mine whether using mobile apps in class had any educational advantage. The results
198 Md A. Hasan et al.
demonstrated one way in which mobile devices, in general, and mobile apps, specif-
ically, can be effectively integrated in an introduction to psychology class to promote
student learning. Their findings suggested that using mobile apps in class could serve
as a useful pedagogical aid to enhance student learning [9].
Apart from these, a vision paper was done for making an impact on developers
from both small and large organizations to develop mobile apps. They mainly focused
on some software development phases (i.e., requirements, security, development,
testing etc.) and discussed about the challenges, future directions and risks of each of
the phases. Their expectance was that the research paper could assist newcomers to
quickly gain a background in the area of mobile apps. Moreover, they hoped that the
argument of the vision for the area will inspire and guide future work and build a
community of researchers with common goals regarding software engineering chal-
lenges for mobile apps [10].
In addition, in a paper, a group of researchers proposed that they had created a
mobile app for the developers to reduce failure of apps as number of failing apps is
growing enormously. They denoted four reasons for failing apps (role of the developer,
user, technical details and marketing) and suggestions to improve them. The initial
motive of this research was to show that the methodological aspect of software engi-
neering can be successfully applied to mobile applications, with the help of software
development life cycle [11]. They used UML, M-UML and other mobile application
development technologies for this purpose.
One more app named “Gate 2015” [12], is available having modules for syllabus,
practice questions, previous question papers, reference of subjects, guidelines etc.
A similar app called “Net Prep+” [13], is available having modules (i.e., Timed test,
Random test, Chapter Wise test etc.). Likewise, another app named “iquiz” was
developed by Kathuria and Gupta that could help students in preparation of competitive
exams like- UGC-NET, GATE etc. Their vision paper also highlighted several chal-
lenges faced by developers in Android App Development [14].
3 Methodology
In this section, we have presented our own approach to find out the specifications for
top mobile applications. For that reason, we have to conduct a survey to collect the
data.
• Smart Phone’s OS
• Internal Storage (8/16/32/64/128/256 GB)
• Smart Phone’s Processor
• List of Apps Installed in the Phone
There were 943 people who had helped us by providing their smart phone’s
information. Among them, 688 people were male and 255 were female. And from our
survey we have found out that most of the smart phone users (827 people) are from 19
to 40 years old. Figure 1 is the overview of our dataset.
4 Our Findings
Figure 4(a) shows Facebook app is mostly used in specification number 26 which is
processor: Quad Core, internal storage: 16 GB and RAM: 1.5 GB. It has appeared 120
times in specification number 26’s phone.
Figure 4(b) shows Instagram app is mostly used in specification number 5 which is
processor: Octa Core, internal storage: 64 GB and RAM: 4 GB and 12 which is
processor: Octa Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 30
times in specification number 5 & 12’s phone.
Figure 4(c) shows Facebook Messenger app is mostly used in specification number
10 which is processor: Octa Core, internal storage: 16 GB and RAM: 2 GB, 5: pro-
cessor: Octa Core, internal storage: 64 GB and RAM: 4 GB and 12: processor: Octa
Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 36, 30 and 30 times in
specification number 10, 5 and 12’s Phone respectively.
Figure 5(a) shows Uber app is mostly used in specification number 5: processor:
Octa Core, internal storage: 64 GB and RAM: 4 GB, 12: processor: Octa Core, internal
storage: 32 GB and RAM: 4 GB and 19: processor: Quad Core, internal storage:
16 GB and RAM: 2 GB. It has appeared 42, 30 & 36 times in specification number 5,
12 & 19’s phone respectively.
Figure 5(b) shows Imo app is mostly used in specification number 10: processor:
Octa Core, internal storage: 16 GB and RAM: 2 GB, 19: processor: Quad Core,
internal storage: 16 GB and RAM: 2 GB and 32: processor: Quad Core, internal
storage: 8 GB and RAM: 2 GB. It has appeared 30, 24 & 18 times in specification
number 10, 19 & 32’s phone respectively.
Figure 5(c) shows Viber app is mostly used in specification number 5: processor:
Octa Core, internal storage: 64 GB and RAM: 4 GB and 18: processor: Octa Core,
internal storage: 64 GB and RAM: 4 GB and 19: processor: Quad Core, internal
storage: 16 GB and RAM: 2 GB. It has appeared 18 times in specification number 5
and 19’s phone.
202 Md A. Hasan et al.
Figure 6(a) shows Snapchat app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB. It has appeared 54 times
in specification number 2’s phone.
Figure 6(b) shows Color Switch app is mostly used in specification number 2:
processor: Octa Core, internal storage: 32 GB and RAM: 3 GB, 20: processor Octa
Core, internal storage: 64 and RAM: 1.5 GB. It has appeared 6 times in specification
number 2 and 20’s phone.
Figure 6(c) shows SHAREit app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB and 10 which is processor:
Octa Core, internal storage: 16 GB and RAM: 2 GB. It has appeared 36 and 18 times
in specification number 2 & 10’s phone.
Figure 7(a) shows Pathao app is mostly used in specification number 9: processor:
Dual Core, internal storage: 64 GB and RAM: 2 GB and 12 which is processor: Octa
Core, internal storage: 32 GB and RAM: 4 GB. It has appeared 18 times in specifi-
cation number 9 & 12’s phone.
Figure 7(b) shows Snapseed app is mostly used in specification number 2: pro-
cessor: Octa Core, internal storage: 32 GB and RAM: 3 GB. It has appeared 42 times
in specification number 2’s phone.
An Analysis on Recent Mobile Application Trend in Bangladesh 203
Figure 7(c) shows Chrome app is mostly used in specification number 2: processor:
Octa Core, internal storage: 32 GB and RAM: 3 GB and 9 which is processor: Dual
Core, internal storage: 64 GB and RAM: 2 GB. It has appeared 18 times in specifi-
cation number 2 and 9’s phone.
5 Conclusion
The given facts and figures should be enough to exhibit the applications generally
being used on smart phones lately. Top-notch industries or high-tech software com-
panies are generally looking for such statistics in order to perceive the types of apps
and specifications on demand. This can save time and unnecessary effort on making
excess apps that are redundant.
App developers and freelancers of our country can also be highly benefited from all
this information given, as it is very important for them to invest their resources in the
right places. Moreover, this methodology can be implemented in other countries. If a
created app does not thrive on the market, such small town businesses or home-based
workshops can suffer extreme loss.
References
1. Eason, G., Noble, B., Sneddon, I.N.: On certain integrals of Lipschitz-Hankel type involving
products of Bessel functions. Philos. Trans. R. Soc. London A247, 529–551 (1955)
2. Clerk Maxwell, J.: A Treatise on Electricity and Magnetism, vol. 2, 3rd edn., pp. 68–73.
Clarendon, Oxford (1892)
3. Jacobs, I.S., Bean, C.P.: Fine particles, thin films and exchange anisotropy. In: Rado, G.T.,
Suhl, H. (eds.) Magnetism, vol. III, pp. 271–350. Academic, New York (1963)
4. Qiu, S., McGee, R.A., Murphey, Y.L.: Adaptive fuzzy prediction for automotive
applications usage. In: 2015 IEEE 14th International Conference on Machine Learning
and Applications (ICMLA) (2015). https://doi.org/10.1109/icmla.2015.138
5. Zhao, F., Shin, J., Reich, J.: Information-driven dynamic sensor collaboration. IEEE Sig.
Process. Mag. 19(2), 61–72 (2002). https://doi.org/10.1109/79.985685
6. Habib, H., Ateeq, M., Bajwa, I.S.: A study on trends of mobile application usage and their
distribution methods in Pakistan. In: Eighth International Conference on Digital Information
Management (ICDIM 2013) (2013). https://doi.org/10.1109/icdim.2013.6693976
7. Acharya, S., Shenoy, A., Lewis, M., Desai, N.: Analysis and prediction of application usage
in android phones. In: 2016 2nd International Conference on Advances in Electrical,
Electronics, Information, Communication and Bio-informatics (AEEICB) (2016). https://doi.
org/10.1109/aeeicb.2016.7538346
8. Feng, Y., Worrachananun, M., Lai, I.K.: Students preferences and intention on using
smartphone education applications. In: 2015 International Symposium on Educational
Technology (ISET) (2015). https://doi.org/10.1109/iset.2015.30
9. Diliberto-Macaluso, K., Hughes, A.: The use of mobile apps to enhance student learning in
introduction to psychology. Teach. Psychol. 43(1), 48–52 (2015). https://doi.org/10.1177/
0098628315620880
10. Nagappan, M., Shihab, E.: Future trends in software engineering research for mobile apps.
In: 2016 IEEE 23rd International Conference on Software Analysis, Evolution, and
Reengineering (SANER) (2016). https://doi.org/10.1109/saner.2016.88
204 Md A. Hasan et al.
11. Inukollu, V.N., Keshamon, D.D., Kang, T., Inukollu, M.: Factors influencing quality of
mobile apps: role of mobile app development life cycle. Int. J. Softw. Eng. Appl. 5(5), 15–34
(2014). https://doi.org/10.5121/ijsea.2014.5502
12. Surachat, K., Kajkamhaeng, S., Damkliang, K., Tiprat, W., Wacharanimit, T.: First aid
application on mobile device. Int. J. Comput. Inf. Eng. 7(5), 622–627 (2013). https://waset.
org/publications/13742/first-aid-application-on-mobile-device. Accessed 20 Apr 2018
13. Joorabchi, M.E., Mesbah, A., Kruchten, P.: Real challenges in mobile app development. In:
2013 ACM/IEEE International Symposium on Empirical Software Engineering and
Measurement (2013). https://doi.org/10.1109/esem.2013.9
14. kathuria, A., Gupta, A.: Challenges in Android application development: a case study. Int.
J. Comput. Sci. Mobile Comput. 4(5), 294–299 (2015). https://www.ijcsmc.com/docs/
papers/May2015/V4I5201557.pdf. Accessed Apr 20
15. Alvi, A.M., Shaon, M.F., Das, P.R., Mustafa, M., Bari, M.R.: Automated course
management system. In: 2017 12th International Conference for Internet Technology and
Secured Transactions (ICITST) (2017). https://doi.org/10.23919/icitst.2017.8356369
Performance Evaluation of Routing
Protocols in DTNs Considering Different
Mobility Models
1 Introduction
Nowdays the number of mobile devices is highly increased. Users can access
information and communicate with each other anywhere and anytime. In high
mobility networks and long distances, end to end connectivity is difficult to
maintain with current technology. In order to deal with this problem, Delay
Tolerant Networks (DTNs) can be used. In DTNs, mobile nodes can send and
receive data, carry data as relays and forward data in opportunistic way upon
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 205–214, 2019.
https://doi.org/10.1007/978-3-030-15035-8_19
206 E. Spaho et al.
contacts. In order to handle disconnections and long delays, DTNs use store-
carry-and-forward approach.
Instead of working end-to-end, in VDTNs, a message-oriented overlay
layer called Bundle layer employs a store-carry-and-forward message switching
paradigm that moves messages from node to node, along a path that eventually
reaches the destination.
In DTNs multiple copy routing protocols are mostly used. Multiple copies
can lead to congestion in networks that are typically characterized by scarce
resources, so a message deletion policy must be used for multi copy routing
schemes. These schemes can use the acknowledgement method to remove the
useless bundles from the network, preventing the nodes from the buffer overflow
problem and avoid transfer of useless message replicas thus relaxing the resources
of the nodes.
Combining routing with a congestion control scheme would manage better
nodes buffer. In this work, we use the congestion control mechanism for DTNs
proposed in [1] and [2]. This mechanism locally advertises the node’s buffer
occupancy to adjacent nodes in order to avoid sending messages to nodes whose
buffers are nearly full. Congestion control mechanism can maximize resource
utilization when resources are available.
In this paper, we investigate and compare the performance of Epidemic,
Spray and Wait and their versions with congestion control and Epidemic with
TCP. Random Waypoint (RWP), Steady State Random Waypoint (SSRWP) and
Tirana map-based mobility models are used. Delivery ratio, hop count, average
delay and average buffer occupancy metrics are used to evaluate the performance.
Simulations are conducted with Network Simulation 3 (ns-3) [3].
The remainder of this paper is as follows. Section 2 introduces DTNs and
routing protocols. The simulation system design and simulation scenarios are
presented in Sect. 3. In Sect. 4 are shown the simulation results. Finally, the
conclusions and future work are presented in Sect. 5.
However, when a big number of copies are created in the network, such approach
increases the contention for network resources as bandwidth and storage.
In [6], authors have studied this model and found that it can provide sub-
stantial capacity at little cost, and that the use of a DTN model often doubles
that capacity compared with a traditional end-to-end model. The main assump-
tion in the Internet that DTNs seek to relax is that an end-to-end path between
a source and a destination exists for the entire duration of a communication
session. When this is not the case, the normal Internet protocols fails. DTNs get
around the lack of end-to-end connectivity with an architecture that is based on
message switching. It is also intended to tolerate links with low reliability and
large delays. The architecture is specified in RFC 4838 [7].
Bundle protocol has been designed as an implementation of the DTN archi-
tecture. A bundle is a basic data unit of the DTN bundle protocol. Each bundle
comprises a sequence of two or more blocks of protocol data, which serve for
various purposes. In poor conditions, bundle protocol works on the application
layer of some number of constituent Internet, forming a store-and-forward over-
lay network to provide its services. The bundle protocol is specified in RFC 5050.
It is responsible for accepting messages from the application and sending them
as one or more bundles via store-carry-and-forward operations to the destination
DTN node. The bundle protocol runs above the TCP/IP level.
already in storage in the other node. The next phase is a check of available
buffer storage space, with the message being forwarded if the other node has
space in its buffer storage. The main goals of this protocol are: maximize the
delivery ratio, minimize the latency and minimize the total resources consumed
in message delivery. It is especially useful when there is lack of information
regarding network topology and nodes mobility patterns.
Spray and Wait Routing Protocol: Spray and Wait [21], is a routing proto-
col that attempts to gain the delivery ratio benefits of replication-based routing
as well as the low resource utilization benefits of forwarding-based routing. The
Spray and Wait protocol is composed of two phases: the spray phase and the wait
phase. When a new message is created in the system, a number L is attached
to that message indicating the maximum allowable copies of the message in the
network. During the spray phase, the source of the message is responsible for
“spraying”, or delivery, one copy to L distinct “relays”. When a relay receives
the copy, it enters the wait phase, where the relay simply holds that particular
message until the destination is encountered directly.
3 Simulation Setup
3.1 Simulation System
The simulations are conducted on ns-3 (ver-3.26) simulation system. ns-3 simu-
lator is a discrete-event network simulator for Internet systems [3]. ns-3 provides
models of how packet data networks work and perform, and provides a simu-
lation engine for users to conduct simulation experiments. Scripting in ns-3 is
done in C++ or Python.
The ns-3 simulator has models for all network elements that comprise a
computer network. For example, network devices represent the physical device
that connects a node to the communication channel. This might be a simple
Ethernet network interface card, or a wireless IEEE 802.11 device.
The ns-3 support standardized output formats for trace data, such as the
pcap format used by network packet analyzing tools. Several external animators
and data analysis and visualization tools can be used with ns-3.
Parameters Values
Number of nodes 100
Simulation time 1800 s
Map size 2.5 km × 2.5 km
Movement Model RWP, SSRWP, Map-based
Antipackets lifetime 10000 s
Antipackets size 10 Bytes
Antipackets header 26 bytes
Congestion threshold 0.7
Bundle lifetime 750 s
Bundle retransmission timeout 1000 s
Number of bundle retransmission 3
Max number of bundle copies for Spray and Wait 16
Hello message interval 100 ms
For all scenarios we evaluate the performance of Epidemic and Spray and
Wait protocols and their versions with congestion control (congestion threshold
0.7) and also use Epidemic with TCP. A bundle can be generated only if the
node has sufficient amount of free buffer space. DTN nodes advertise their buffer
content to each other every 100 ms by sending Hello messages. Moreover, if DTN
congestion control is enabled, a bundle can be forwarded only to such interme-
diate nodes that have enough free buffer space. Simulation parameters and their
values are shown in Table 1.
In following, we will briefly describe different mobility models considered in
this work. The RWP model is a commonly used mobility model for simulations
of wireless communication networks. Basically, in RWP every node picks up a
random destination and a random velocity at certain points called waypoints.
This model has been extended in a number of ways in order to take into account
more realistic movements.
SSRWP model [24,25] is based on RWP model for case when speed, pause
and position are uniformly distributed random variables. The difference is that
the initial values of these parameters are not from uniform distribution but from
stationary distribution of RWM model.
Tirana map-based mobility model is a realistic scenario with vehicles moving
in Tirana city created by using Open Street Map (OSM) [26] and Simulation of
Urban MObility (SUMO) [27] (see Figs. 1 and 2).
4 Simulation Results
The simulation results of delivery ratio for RWP, SSRWP and Tirana map-
based mobility model are shown in Fig. 3. Delivery ratio is calculated as the
210 E. Spaho et al.
1
Binary Spray and Wait
Binary Spray and Wait+CC and Tc=0.7
Epidemic
Epidemic+CC and Tc=0.7
Epidemic+TCP
0.8
Delivery ratio
0.6
0.4
0.2
0
RWP SSRWP Tirana Map-based
Mobility Model
600
Binary Spray and Wait
Binary Spray and Wait+CC and Tc=0.7
Epidemic
Epidemic+CC and Tc=0.7
Epidemic+TCP
500
400
Delay (s)
300
200
100
0
RWP SSRWP Tirana Map-based
Mobility Model
5
Binary Spray and Wait
Binary Spray and Wait+CC and Tc=0.7
Epidemic
Epidemic+CC and Tc=0.7
Epidemic+TCP
4
Hop number
0
RWP SSRWP Tirana Map-based
Mobility Model
Buffer occupancy
6
2x10
1.5x106
1x106
500000
0
RWP SSRWP Tirana Map-based
Mobility Model
shows that SSRWP performs better than RWP. For Spray and Wait and Epi-
demic with and without congestion control the delays are almost the same. For
Epidemic with TCP, the delay is higher because of acks.
The results for hop count are shown in Fig. 5. Hop count is the average
number of hops counts between the source and destination. Hop count is almost
the same for both scenarios because the number of nodes is the same. Epidemic
protocol will have higher hop count because it uses more network resources
(intermediate nodes) compared with Spray and Wait.
In Fig. 6 are shown the results for average buffer occupancy. Average buffer
occupancy shows the average amount of bytes stored in the buffer nodes. For
SSRWP, we can see a decrease of buffer occupancy for all protocols. If we have
a node “a” that have met before node “b” than it has more chances to meet it
again compared with RWP. In the encounters after the first time, nodes exchange
also antipackets that delete duplicated bundles from the buffers. The importance
of congestion control is that it controls the buffers from being full. The usage of
congestion control and the good management of buffers leads to better delivery
ratio. For Epidemic with TCP, the buffer occupancy is higher than Spray and
Wait because more packets are created.
5 Conclusions
• The protocols performance using SSRWP and RWP is higher compared with
the realistic scenarios because in the realistic scenario the nodes movement is
more complex and the opportunistic contacts are more rare.
Performance Evaluation of Routing Protocols in DTNs 213
• When SSRWP is used the network will be more stable, more contacts between
nodes can occur and more bundles can be delivered compared with RWP.
• Congestion control improves the delivery ratio and buffer management for
both protocols. It uses a mechanism to stop bundle generation and save buffers
from being full.
• The usage of Epidemic with TCP leads to high average delay in all scenarios
because of ack packets.
References
1. Lakkakorpi, J., Ginzboorg, P.: ns-3 module for routing and congestion control
studies in mobile opportunistic DTNs. In: Proceedings of Performance Evaluation
of Computer and Telecommunication Systems (SPECTS), pp. 1–5 (2013)
2. Lakkakorpi, J., Pitkanen, M., Ott, J.: Using buffer space advertisements to avoid
congestion in mobile opportunistic DTNs. In: Proceedings of WWIC, Barcelona,
Spain, pp. 1–12, June 2011
3. The ns-3 Network Simulator. http://www.nsnam.org
4. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Pro-
ceedings of the International Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, ser. SIGCOMM 2003, pp.
27–34 (2003)
5. Delay- and disruption-tolerant networks (DTNs) tutorial, NASA/JPL’s Inter-
planetary Internet (IPN) Project (2012). http://www.warthman.com/images/
DTNTutorialv2.0.pdf
6. Laoutaris, N., Smaragdakis, G., Rodriguez, P., Sundaram, R.: Delay tolerant bulk
data transfers on the Internet. In: Proceedings of the 11th International Joint
Conference on Measurement and Modeling of Computer Systems (SIGMETRICS
2009), pp. 229–238 (2009)
7. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss,
H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April
2007
8. Massri, K., Vernata, A., Vitaletti, A.: Routing protocols for delay tolerant net-
works: a quantitative evaluation. In: Proceedings of ACM workshop PM2HW2N
2012, pp. 107–114 (2012)
9. Massri, K., Vitaletti, A., Vernata, A., Chatzigiannakis, I.: Routing protocols for
delay tolerant networks: a reference architecture and a thorough quantitative
evaluation. J. Sens. Actuator Networks, 1–28 (2016). https://doi.org/10.3390/
jsan5020006
10. Demmer, M., Fall, K.: DTLSR: delay tolerant routing for developing regions. In:
Proceedings of the 2007 ACM Workshop on Networked Systems for Developing
Regions, 6 p. (2007)
11. Ilham, A.A., Niswar, M., Agussalim: Evaluated and optimized of routing model on
Delay Tolerant Network (DTN) for data transmission to remote area. In: Proceed-
ings of FORTEI, Indonesia University Jakarta, pp. 24–28 (2012)
214 E. Spaho et al.
12. Uchida, N., Ishida, T., Shibata, Y.: Delay tolerant networks-based vehicle-to-
vehicle wireless networks for road surveillance systems in local areas. Int. J. Space-
Based Situated Comput. 6(1), 12–20 (2016)
13. Bylykbashi, K., Spaho, E., Barolli, L., Xhafa, F.: Routing in a many-to-one com-
munication scenario in a realistic VDTN. J. High Speed Networks 24(2), 107–118
(2018)
14. Bylykbashi, K., Spaho, E., Barolli, L., Xhafa, F.: Impact of node density and TTL
in vehicular delay tolerant networks: performance comparison of different routing
protocols. Int. J. Grid Util. Comput. 7(3), 136–144 (2017)
15. Jain, S., Fall, K., Patra, R.: Routing in a delay tolerant network. In: Proceedings of
ACM SIGCOMM 2004 Conference on Applications, Technologies, Architectures,
and Protocols for Computer Communication, Portland, Oregon, USA, 30 August–3
September 2004, pp. 145–158 (2004)
16. Zhang, Z.: Routing in intermittently connected mobile ad hoc networks and delay.
IEEE Commun. Surv. Tutorials 8(1), 24–37 (2006)
17. Soares, V.N.G.J., Rodrigues, J.J.P.C., Farahmand, F.: GeoSpray: a geographic
routing protocol for vehicular delay-tolerant networks. Inf. Fusion 15(1), 102–113
(2014)
18. Burgess, J., Gallagher, B., Jensen, D., Levine, B.N.: Maxprop: routing for vehicle-
based disruption-tolerant networks. In: Proceedings of the IEEE Infocom, April
2006
19. Lindgren, A., Doria, A., Davies, E., Grasic, S.: Probabilistic routing protocol for
intermittently connected networks. draft-irtf-dtnrg-prophet-09. http://tools.ietf.
org/html/draft-irtf-dtnrg-prophet-09
20. Vahdat, A., Becker, D.: Epidemic routing for partially connected ad hoc networks.
Technical report CS-200006, Duke University, April 2000
21. Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and Wait: an efficient
routing scheme for intermittently connected mobile networks. In: Proceedings of
ACM SIGCOMM 2005 - Workshop on Delay Tolerant Networking and Related
Networks (WDTN-05), Philadelphia, PA, USA, pp. 252–259 (2005)
22. Bylykbashi, K., Spaho, E., Barolli, L., Takizawa, M.: Comparison of spray and
wait and epidemic protocols in different DTN scenarios. In: Proceedings of the
12th International Conference on Broad-Band Wireless Computing, Communica-
tion and Applications (BWCCA-2017), pp. 218–229 (2017)
23. Spaho, E., Bylykbashi, K., Barolli, L., Takizawa, L.: Routing in a DTN: perfor-
mance evaluation for random waypoint and steady state random waypoint using
NS3 simulator. In: Proceedings of the 12th International Conference on P2P, Par-
allel, Grid, Cloud and Internet Computing (3PGCIC-2017), pp. 133–141 (2017)
24. Navidi, W., Camp, T.: Stationary distributions for the random waypoint mobility
model. IEEE Trans. Mob. Comput. 3(1), 99–108 (2004)
25. Navidi, W., Camp, T., Bauer, N.: Improving the accuracy of random waypoint
simulations through steady-state initialization. In: Proceedings of the 15th Inter-
national Conference on Modeling and Simulation (MS 2004), pp. 319–326, March
2004
26. Open street map. http://www.openstreetmap.org/
27. Simulation of urban mobility. http://sumo.sourceforge.net/
A Distance-Based
Advertisement-Delivery Method
for Vehicular DTN
1 Introduction
Recently, many vehicle companies have been developing vehicles together with
ICT companies to provide the added value of services. Traditionally, vehicles
have been used for people moving and baggage transfer, but until now vehicles
are going to be connected to the network and becoming mobile terminals [4,7,12].
Vehicles will connect to various wireless networks that will become relay nodes
and will have various applications [9,13,21,23,27]. The vehicles will be equipped
with On-Board-Unit (OBU) by plug-in. However, there are many problems that
used to be solved.
Delay/Disruption/Disconnection Tolerant Networking (DTN) is a message
relaying method in Vehicular to Vehicular (V2V) communication [11,17,19,20,
22]. In Vehicular DTN, consumption of network resources and storage usage of
each node becomes a critical problem due to the DTN nodes duplicate messages
to other nodes.
In our previous work [10], we have proposed an Enhanced Message Suppres-
sion Controller (EMSC) for Vehicular-DTN. We have shown that the EMSC
can reduce the consumption of network resources. But, EMSC needs some addi-
tional control messages to reduce the replicated bundle messages. In addition,
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 215–223, 2019.
https://doi.org/10.1007/978-3-030-15035-8_20
216 S. Nakasaki et al.
2 Related Work
DTN can provide a reliable internet-working for space tasks [1,3,5,8,15,25].
The space networks have possibly long delay, frequent link disconnection and
frequent disruption. In DTN, the messages are stored and forwarded by nodes.
When the nodes receive messages, they store the messages in their storage. After
that, the nodes duplicate the messages to other nodes when it is transmitted.
This technique is called message switching. The architecture is specified in RFC
4838 [6].
Epidemic routing is well-known routing protocol for DTN [14,16,24]. Epi-
demic routing uses two control messages to duplicate messages. Nodes periodi-
cally broadcast Summary Vector (SV) message in the network. The SV contains
a list of stored messages of each node. When the nodes receive the SV, they com-
pare received SV to their SV. The nodes send REQUEST message if received
SV contains unknown messages.
In Epidemic routing, consumption of network resources and storage usage
become a critical problem. Because the nodes duplicate messages to neighbors
in its communication range. Moreover, received messages remain in the stor-
age, because the messages are continuously duplicated even if destination node
receives the messages. Therefore, recovery schemes such as timer or anti-packet
are needed to limit the duplicate messages. In case of the timer, messages have
lifetime. The messages are punctually deleted when the lifetime of the mes-
sages is expired. However, setting of a suitable lifetime is difficult. In case of
anti-packet, destination node broadcasts the anti-packet, which contains the list
of messages that are received by the destination node. Nodes delete the mes-
sages according to the anti-packet. Then, the nodes duplicate the anti-packet to
other nodes. However, network resources are consumed by the anti-packet. In
this paper, we propose a distance-based message relaying method to limit the
duplicate messages in Vehicular DTN.
We use two types of routing protocols and send bundle messages from starting-
point (car shop) to all vehicles considering ITU-R P.1411 propagation model [2].
Therefore, we consider interference of obstacles on 5.9 GHz band.
1. Conventional Epidemic
2. Epidemic with proposed distance-based message relaying method
Parameter Value
Simulation time (Tmax ) 600 [s]
Area dimensions 2, 000 [m] × 2, 000 [m]
Number of vehicles 100, 200, 300 [vehicles]
Service range 1, 000 [m]
Minimum speed (Vmin ) 8.333 [m/s]
Maximum speed (Vmax ) 16.666 [m/s]
PHY model IEEE 802.11p
Propagation model ITU-R P.1411
Antenna model Omni-directional
We consider a grid scenario with 100, 200 and 300 vehicles. We present the sim-
ulation scenario for 100 vehicles in Fig. 3. Table 1 shows the simulation parame-
ters. Message starting point is static, and other vehicles move on road based on
random way-point mobility model.
For simulations, we consider two evaluation parameters: position distribution
and delay. The position distribution indicates the vehicle positions when vehicle
receives the bundle message in service range. The delay indicates the duration
of the message to reach the end of service range. We show the edge nodes on
simulation area in Fig. 3.
Fig. 4. Position distribution, when vehicles receive the bundle messages using conven-
tional Epidemic.
220 S. Nakasaki et al.
Fig. 5. Position distributions for different vehicles, when vehicles receive the bundle
messages using proposed method. Red line indicates the edge of service range.
A Distance-Based Advertisement-Delivery Method for Vehicular DTN 221
Table 2. Results of delay for different vehicles, which are measured by edge of service
range.
5 Conclusions
In this paper, we proposed a distance-based recovery method for Vehicular DTN.
We evaluated the proposed recovery method considering position distribution
and delay as evaluation metrics. From the simulation results, we observed that
our proposed method limits the replicated bundle messages considering specified
area by applying the distance-based advertisement-delivery method.
In the future work, we would like to add new functions and make extensive
simulations considering different parameters.
References
1. Delay- and disruption-tolerant networks (DTNs) tutorial. NASA/JPL’s Interplan-
etary Internet (IPN) Project (2012). http://www.warthman.com/images/DTN
Tutorial v2.0.pdf
2. Recommendation ITU-R P.1411-7: Propagation data and prediction methods for
the planning of short-range outdoor radiocommunication systems and radio local
area networks in the frequency range 300 MHz to 100 GHz. ITU (2013)
3. Araniti, G., Bezirgiannidis, N., Birrane, E., Bisio, I., Burleigh, S., Caini, C., Feld-
mann, M., Marchese, M., Segui, J., Suzuki, K.: Contact graph routing in DTN
space networks: overview, enhancements and performance. IEEE Commun. Mag.
53(3), 38–46 (2015)
4. Araniti, G., Campolo, C., Condoluci, M., Iera, A., Molinaro, A.: LTE for vehicular
networking: a survey. IEEE Commun. Mag. 21(5), 148–157 (2013)
5. Caini, C., Cruickshank, H., Farrell, S., Marchese, M.: Delay- and disruption-
tolerant networking (DTN): an alternative solution for future satellite networking
applications. Proc. IEEE 99(11), 1980–1997 (2011)
222 S. Nakasaki et al.
6. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fall, K., Weiss,
H.: Delay-tolerant networking architecture. IETF RFC 4838 (Informational), April
2007
7. Dias, J.A.F.F., Rodrigues, J.J.P.C., Xia, F., Mavromoustakis, C.X.: A cooperative
watchdog system to detect misbehavior nodes in vehicular delay-tolerant networks.
IEEE Trans. Ind. Electron. 62(12), 7929–7937 (2015)
8. Fall, K.: A delay-tolerant network architecture for challenged Internets. In: Pro-
ceedings of the International Conference on Applications, Technologies, Architec-
tures, and Protocols for Computer Communications, SIGCOMM 2003, pp. 27–34
(2003)
9. Grassi, G., Pesavento, D., Pau, G., Vuyyuru, R., Wakikawa, R., Zhang, L.: VANET
via named data networking. In: Proceedings of the IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS 2014), pp. 410–415, April 2014
10. Ikeda, M., Ishikawa, S., Barolli, L.: An enhanced message suppression controller for
vehicular-delay tolerant networks. In: Proceedings of the 30th IEEE International
Conference on Advanced Information Networking and Applications (IEEE AINA-
2016), pp. 573–579, March 2016
11. Mahmoud, A., Noureldin, A., Hassanein, H.S.: VANETs positioning in urban envi-
ronments: a novel cooperative approach. In: Proceedings of the IEEE 82nd Vehic-
ular Technology Conference (VTC-2015 Fall), pp. 1–7, September 2015
12. Ohn-Bar, E., Trivedi, M.M.: Learning to detect vehicles by clustering appearance
patterns. IEEE Trans. Intell. Transp. Syst. 16(5), 2511–2521 (2015)
13. Radenkovic, M., Walker, A.: CognitiveCharge: disconnection tolerant adaptive col-
laborative and predictive vehicular charging. In: Proceedings of the 4th ACM
MobiHoc Workshop on Experiences with the Design and Implementation of Smart
Objects (SMARTOBJECTS-2018), June 2018
14. Ramanathan, R., Hansen, R., Basu, P., Hain, R.R., Krishnan, R.: Prioritized epi-
demic routing for opportunistic networks. In: Proceedings of the 1st International
MobiSys Workshop on Mobile Opportunistic Networking (MobiOpp 2007), pp.
62–66 (2007)
15. Rüsch, S., Schürmann, D., Kapitza, R., Wolf, L.: Forward secure delay-tolerant
networking. In: Proceedings of the 12th Workshop on Challenged Networks
(CHANTS-2017), pp. 7–12, October 2017
16. Rohrer, J.P., Mauldin, A.N.: Implementation of epidemic routing with IP conver-
gence layer in ns-3. In: Proceedings of the 10th Workshop on NS-3 (WNS3-2018),
pp. 69–76, June 2018
17. Sato, F., Kikuchi, R.: Hybrid routing scheme combining with geo-routing and DTN
in VANET. In: Proceedings of the 10th International Conference on Innovative
Mobile and Internet Services in Ubiquitous Computing (IMIS-2016), pp. 250–255,
July 2016
18. Scenargie: Space-time engineering, LLC. http://www.spacetime-eng.com/
19. Stute, M., Maass, M., Schons, T., Hollick, M.: Reverse engineering human mobility
in large-scale natural disasters. In: Proceedings of the 20th ACM International
Conference on Modelling, Analysis and Simulation of Wireless and Mobile Systems
(MSWiM-2017), pp. 219–226, November 2017
20. Theodoropoulos, T., Damousis, Y., Amditis, A.: A load balancing control algorithm
for EV static and dynamic wireless charging. In: Proceedings of the IEEE 81st
Vehicular Technology Conference (VTC-2015 Spring), pp. 1–5, May 2015
21. Tornell, S.M., Calafate, C.T., Cano, J.C., Manzoni, P.: DTN protocols for vehicular
networks: an application oriented overview. IEEE Commun. Surv. Tutor. 17(2),
868–887 (2015)
A Distance-Based Advertisement-Delivery Method for Vehicular DTN 223
22. Uchida, N., Ishida, T., Shibata, Y.: Delay tolerant networks-based vehicle-to-
vehicle wireless networks for road surveillance systems in local areas. Int. J. Space
Based Situated Comput. 6(1), 12–20 (2016)
23. Urquiza-Aguiar, L., Igartua, M.A., Tripp-Barba, C., Calderón-Hinojosa, X.:
2hGAR: 2-hops geographical anycast routing protocol for vehicle-to-infrastructure
communications. In: Proceedings of the 15th ACM International Symposium
on Mobility Management and Wireless Access (MobiWac-2017), pp. 145–152,
November 2017
24. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks.
Duke University, Technical report (2000)
25. Wyatt, J., Burleigh, S., Jones, R., Torgerson, L., Wissler, S.: Disruption toler-
ant networking flight validation experiment on NASA’s EPOXI mission. In: Pro-
ceedings of the 1st International Conference on Advances in Satellite and Space
Communications (SPACOMM-2009), pp. 187–196, July 2009
26. Yoshino, Y., Nakasaki, S., Ikeda, M., Barolli, L.: A threshold-based adaptive
method for message suppression controller in vehicular DTNs. In: Proceedings
of the 13th International Conference on Broad-Band Wireless Computing, Com-
munication and Applications (BWCCA-2018), pp. 517–524, October 2018
27. Zhang, W., Jiang, S., Zhu, X., Wang, Y.: Cooperative downloading with privacy
preservation and access control for value-added services in VANETs. Int. J. Grid
Util. Comput. 7(1), 50–60 (2016)
Numerical Simulation of Glide Slope
Signal Interferences by Irregular Ground
1 Introduction
Instrument Landing System (ILS) plays an important role in the safe landing
of aircraft at airports. ILS consists of a Localizer (LOC), Glide Slope (GS), and
Maker [1]. The GS provides guidance on the vertical position of the aircraft
with reflect to the approach path to the runway center [1]. The radiation from
GS antennas forms an approach path using ground-reflections. Specification of
GS is evaluated using the difference of the depth of the modulation (DDM).
When the aircraft moves along an appropriate approach path, DDM becomes
zero, resulting in no movement of the cross point indicator [2]. The value of
the DDM is regulated by the International Civil Aviation Organization (ICAO)
[1], and the quality of the GS is classified into three categories (CAT I, CAT
II, and CAT III). The GS specification of each airport environment should be
satisfied with the determined category. However, if there are irregular ground
features around the GS antenna unnecessary scattered waves are formed, which
impact the approach path. Therefore, it is extremely important to investigate
such multipath environment to maintain the quality of the GS.
Presently, there are many studies on ILS signals [3–11]. In most studies,
ground model for numerical simulation was limited to two-dimensional (2D)
situation, which changes only area in front of GS antenna. However, in reality,
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 224–233, 2019.
https://doi.org/10.1007/978-3-030-15035-8_21
Numerical Simulation of Glide Slope Signal Interferences 225
there are often slopes on the ground, which have a former area and side area,
resulting in a three dimensional (3D) ground situation. As far as we know, current
comparisons between 2D and 3D models have been insufficient. The purpose of
this study is to analyze the multipath interference caused by irregular ground.
A numerical method based on the ray-tracing method (RTM) [12–14] is used
to compute electromagnetic fields above 2D and 3D ground models. The RTM
is an approximate solution that does not need a lot of computational memory.
Therefore, it can treat a large propagation environment. Here, irregular ground
models are numerically generated by Gaussian random valuable installed to the
software. By comparing the results of 2D and 3D ground models, we evaluate
how the proposed numerical method copes with the different ground shapes.
2 Glide Slope
GS guides an aircraft regarding its vertical position with respect to the app-
roach path. Figure 1 illustrates an approach path formed by the radiation of
GS antenna. In Japan, an M-array transmission system is employed using a
frequency band in the range 329–335 MHz, where the transmitted signals are
generated by amplitude modulation. There are three components to GS radia-
tion: carrier, sideband and clearance. The specification of the GS is defined by
the DDM, which is computed by the difference of the depth of the modulation
between 90 and 150 Hz [1]. GS is located as shown in Fig. 1b which becomes
generally the touch down point. In this figure, L is the distance from the end of
the runway, and D is the distance from the center of the runway.
An approach path is generated from direct waves and ground-reflection
waves. Ground reflections are observed from flat ground in the basic concept
of GS. Scattered waves result in the course disturbance when the ground surface
in front of GS antenna is irregular. Figure 2 illustrates an overview of the elec-
tromagnetic wave interference. Each airport manages an operation method that
suites the category (CAT I, II, or III), and the DDM should be satisfied by the
category in operation. To satisfy the operation category, a numerical simulation
would be an effective solution.
3 Numerical Method
Most research papers on GS considered the ground to be a 2D model, which
changes the height of only former region of the antenna [10,11]. However, in
reality, the changes are in the former region and the side region, therefore, 3D
ground model must be assumed. The behaviors of electromagnetic waves above
a 3D ground model might be more complicated than that of a 2D model. Here,
we propose a numerical method for a 3D ground model and compare its results
to those from a 2D model.
In this paper, RTM is introduced to analyze electromagnetic waves above
irregular ground. It can treat a relatively large analytical region than other
methods [15,16]. First, the RTM searches for rays from a source to a receiver.
226 J. Honda et al.
Second, the RTM computes electromagnetic fields based on the ray data. The
computational time required to search for the rays depends on the propagation
environment. In general, the searching rays dominate about 90% of the total
computation time. Since our target is a large analytical region, a simplified RTM
Numerical Simulation of Glide Slope Signal Interferences 227
(b) 3D ground model in view of x-y plane. The height in z-axis is given at each intersection between
triangular faces.
model than a 2D model. However, the 3D ground model is much closer to the
real environment.
This study considered the shortest path along a straight line between a source to
a receiver. Any other plate/face that is not in a straight line is ignored to simplify
the computation. The first step is to evaluate the relation between a source, a
receiver, and each face. The center position on each face is then defined. If they
are in line of sight (LOS), an approximate ray is generated from the source to
the receiver via the central position of each face. This should be modified for
real rays that satisfy Snell’s law using the imaging method. Then, all estimated
rays are filtered to avoid the transmission of the rays against other faces. These
modified rays are reflection rays. Conversely, diffracted rays are estimated at all
faces, which become LOS with the source and receiver. Here, we assume that
only diffracted waves along a straight line between the source and the receiver
is dominant in the total field. Actually, all regions of irregular ground cause
diffraction and reflection, but here we ignore these interferences. Figure 4 shows
a schematic of ray searching above a 2D ground model. The basic ray searching
of a 3D ground model is similar to the 2D ground model.
Fig. 4. The principle of search rays that are connected by the shortest path along a
straight line from source to receiver.
Numerical Simulation of Glide Slope Signal Interferences 229
Eru,m,l
= ⎣ i
(Dnm ) · s u,m,l
(Dnk ) · E0 ⎦ (1)
n=1 m=1 k=1 rnu,m,l
where superscripts u, m, and l represent the upper, middle, and lower antennas,
respectively. N is the total number of rays, Mni is the number of times of image
diffraction, and Mns is the number of times of source diffraction. Ds and Di are
the diffraction coefficients for the source and image, respectively. The reflection
coefficient is included in the image diffraction. The distance of the n-th ray from
the source to the receiver is given by:
i s
k=Mn
+Mn
rnu,m,l = rnk (n = 1, 2, . . . , N ) (2)
k=0
where rnk is the k-th distance from one reflection or diffraction point to the
next one. Diffraction is divided into two parts, i.e., source diffraction and image
diffraction [15], and they work to keep the field continuity as discussed in detail
elsewhere [16].
4 Numerical Result
Figure 5 shows the generated 2D and 3D ground models. Ground shapes were
give by Gaussian random valuable installing in the software. Maximum height
is about 2.5 m, and minimum height is about −3.0 m. The ground in the 2D
model changes in the direction of the forward area, whereas the ground in the
3D model is generated from the forward area and side area. Therefore, the 3D
ground model is more complicated than the 2D model. The GS antenna is located
at L = 315 m and D = 120 m.
Figure 6 shows some examples of search rays, including the first reflection and
diffraction. The receiver is located at x = 1000 m, and the height is computed
to be 3.0◦ of approach path. In Fig. 6a, direct rays and diffracted rays caused by
the edge of the ground appear. Rays are emitted from three antennas installed
to the GS. In Fig. 6b, direct, reflected, and diffracted rays are computed. It is
shown that the propagation path above 3D ground is more complicated than
that above 2D ground. In the 3D case, the rays are estimated along a straight
line between the source and the receiver where many intersections exist.
Based on previous ray information, we can compute the DDMs for 2D and
3D models. Figure 7 shows a comparison of DDMs, which are computed above
flat, 2D, and 3D grounds. The red line is the DDM above flat ground. This
expresses the ideal condition where no interference is detected. The green line
shows the DDM above 2D ground, and some small interference is observed, which
230 J. Honda et al.
gradually increases when the receiver is far from the GS antenna. This might be
caused by an average slope in front of the antenna, and is regarded as a primary
plane. The blue line shows the DDM above 3D ground. A lot of interference is
observed because the ground is more complicated than that in the 2D model.
There is a discontinuous section between 3500 m and 6200 m. In this region, a
reflected wave that satisfied Snell’s law is computed, but reflections in the other
regions cannot be detected, except for the near field. In this simulation, ground
Numerical Simulation of Glide Slope Signal Interferences 231
5 Conclusion
References
1. ICAO (International Civil Aviation Organization): Aeronautical Telecommunica-
tions, Annex 10, vol. 1, 6th edn. (2006)
2. Chin, G., Jordan, L., Kahn, D., Morin, S.: Instrument landing system performance
prediction. IEEE MTT-S International Microwave Symposium, pp. 346–348, May
1975
3. Redlich, R.W., Gorman, J.T.: Disturbance of ILS localizer signals by reflections
from large hangers. IEEE Trans. Aerosp. Electron. Sys. AES-5, 1001–1002 (1969)
Numerical Simulation of Glide Slope Signal Interferences 233
4. Shin, S.L.: ILS localizer multipath analysis. IEEE Trans. Aerosp. Electron. Sys.
AES-7(1), 54–60 (1971)
5. Thain, A., Estienne, J.-P., Robert, J., Peres, G., Cambon, G., Evain, L., Splitz, B.:
A solution for ILS disturbance due to a building. In: Proceedings of the 6th Euro-
pean Conference on Antenna and Propagation (EuCAP), pp. 2392–2395, March
2012
6. Thain, A., Estienne, J.P., Peres, G., Spitz, B., Evain, L., Kumar, M.: Comparisons
of different approaches for ILS simulation. In: Proceedings of the 4th European
Conference on Antenna and Propagation (EuCAP), Barocelona, Spain, pp. 1–5,
April 2010
7. Geise, R., Schueuer, J., Thiele, L., Notte, K., Beckers, T., Enders, A.: A slotted
waveguide setup as scaled instrument-landing-system for measuring scattering of
an A380 and large objects. In: Proceedings of the Fourth European Conference on
Antennas and Propagation (EuCAP 2010), Barcelona, Spain, pp. 227–231, April
2010
8. Greving, G., Biermann, W.D.: Theoretical aspect and numerical solution of scat-
tered from objects under grazing angle incidence - the A380 and the Aircraft Tailfin.
In: Proceedings of the 37th European Microwave Conference, Munich, Germany,
pp. 1369–1372, October 2007
9. Godfrey, J., et al.: Terrain modeling using the half-plane geometry with appli-
cations to ILS glide slope antenna. IEEE Trans. Antennas Propag. 24, 370–378
(1976)
10. Luebbers, R., Ungvichian, V., Mitchell, L.: GTD terrain reflection model applied
to ILS glide scope. IEEE Trans. Aerosp. Electron. Sys. AES-18, 11–20 (1982)
11. Walton, E.K.: Effect of wet snow on the null-reference ILS system. IEEE Trans.
Aerosp. Electron. Syst. 29, 1030–1035 (1993)
12. Honda, J., Otsuyama, T.: An estimation algorithm of scattered powers caused by
a moving aircraft. IEICE Commun. Express 2(11), 490–495 (2013)
13. Honda, J., Otsuyama, T.: Estimation of target detection rate in aircraft surveillance
system. In: Proceedings of the 17th International Conference on Network-Based
Information Systems (NBiS 2014), Salerno, Italy, pp. 561–565, September 2014
14. Honda, J., Otsuyama, T.: Rapid computation algorithm for radio wave propaga-
tion characteristics on airport surface. In: Proceedings of the Eighth International
Conference on Complex Intelligent and Software Intensive Systems (CISIS 2014),
Birmingham, United Kingdom, pp. 302–306, July 2014
15. Honda, J., Uchida, K., Yoon, K.Y.: Estimation of radio communication distance
along random rough surface. IEICE Trans. ELECTRON. E93-C(1), 39–45 (2010)
16. Takematsu, M., Uchida, K., Honda, J.: Method for reduction of field computation
time for discrete ray tracing method. IEICE Trans. Electron. E97-C(3), 198–206
(2014)
Effect of Client Priority
in the Performance of a Fuzzy-Based
WLAN Triage System
1 Introduction
With the development of wireless technology and Internet, there is an increasing
need towards portable and mobile computers such as smart phones [1,24]. The
wireless networks need to provide communications between mobile terminals.
The Wireless Local Area Networks (WLANs) provide high bandwidth access for
users in a limited geographical area. With the popularization of mobile devices,
many device communicate together over WLANs [27]. WLANs have a lot of
restriction on communication resources comparing wired LANs. Therefore, it is
more difficult to guarantee the Quality of Service (QoS).
In our previous work, the admission decision is done by a Fuzzy-based Admis-
sion Control System (FACS) [7–9]. In this paper, we present the effect of client
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 234–243, 2019.
https://doi.org/10.1007/978-3-030-15035-8_22
Effect of Client Priority 235
Fig. 1. CSMA/CA.
2 CSMA/CA
WLAN is standardized by IEEE 802.11 which uses CSMA/CA as shown in
Fig. 1. The nodes check whether other devices are communicating or not before
starting communication [5]. If other nodes are communicating with the AP, the
node waits for a period of time, called Distributed Inter Frame Space (DIFS).
After that, it waits an additionally random time called back-off time. After the
back-off time, if other devices are not communicating, the node starts to send
data [25]. AP which received the data waits a constant time, called Short Inter
Frame Space (SIFS) and sends ACK to the node which sent the data. Because
ACK frame should be sent soon, it is shorter than DIFS.
The IEEE 802.11e standard is an important extension of the IEEE 802.11
standard focusing on QoS [16] that works with any PHY implementation. Wire-
less nodes equipped with IEEE 802.11e features are now known as QoS stations
(QSTAs) and they are associated with a QoS access point (QAP) to form a QoS
basic service set (QBSS). The main feature of the IEEE 802.11e standard is
that it improves the MAC layer for QoS provisioning by providing support for:
segregation of data packets based on priority requirements; negotiation of QoS
parameters through a central coordinator or AP; and admission control.
The IEEE 802.11e standard introduces EDCF. This scheme is useful for QoS
provisioning to support delay-sensitive voice and video applications [2].
In the DCF configuration, a contention window is set after a frame is trans-
mitted. This is done to avoid any collisions. The window defines the contention
time of various stations who contend with each other for access to channel. How-
ever, each of the stations cannot size the channel immediately, rather the MAC
protocol uses a randomly chosen time period for each station after that channel
has undergone transmission [28].
236 K. Ozera et al.
The EDCF uses this contention window to differentiate between high prior-
ity and low priority services [23]. The central coordinator assigns a contention
window of shorter length to the stations with higher priority that helps them to
transmit before the lower priority ones [22]. To differentiate further, Inter Frame
Spacing (IFS) can be varied according to different traffic categories. Instead of
using a DIFS as for the DCF traffic, a new inter-frame spacing called Arbitration
Inter Frame Spacing (AIFS) is used. The AIFS used for traffic has a duration
of a few time slots longer than the DIFS duration. Therefore, a traffic category
having smaller AIFS gets higher priority.
IEEE802.11e provides EDCF and Hybrid coordination function controlled
channel access (HCCA) as a priority control method [2,17,30]. Mainly, EDCF
is used because of easy implementation and compatibility of CSMA/CA.
3.1 FC
In the FC systems, expert knowledge is encoded in the form of fuzzy rules, which
describe recommended actions for different classes of situations represented by
fuzzy sets.
In fact, any kind of control law can be modeled by the FC methodology,
provided that this law is expressible in terms of “if ... then ...” rules, just like
in the case of expert systems. However, FL diverges from the standard expert
system approach by providing an interpolation mechanism from several rules. In
the contents of complex processes, it may turn out to be more practical to get
knowledge from an expert operator than to calculate an optimal control, due to
modeling costs or because a model is out of reach.
3.3 FC Rules
FC describes the algorithm for process control as a fuzzy relation between infor-
mation about the conditions of the process to be controlled, x and y, and the
output for the process z. The control algorithm is given in “if ... then ...” expres-
sion, such as:
If x is small and y is big, then z is medium;
If x is big and y is medium, then z is big.
These rules are called FC rules. The “if” clause of the rules is called the
antecedent and the “then” clause is called consequent. In general, variables x
and y are called the input and z the output. The “small” and “big” are fuzzy
values for x and y, and they are expressed by fuzzy sets.
Fuzzy controllers are constructed of groups of these FC rules, and when an
actual input is given, the output is calculated by means of fuzzy inference.
The defuzzification operation produces a non-FC action that best represent the
membership function of an inferred FC action. Several defuzzification methods
have been suggested in literature. Among them, four methods which have been
applied most often are:
4 Proposed System
The proposed system model is show in Fig. 2. For implementing our proposed sys-
tem, we consider 4 linguistic input parameters: Connected Keeping Time (CKT),
User Priority (UP), Received Signal Strength Indication (RSSI) and Number of
Packet (NP) to decide the Connection Priority (CP) output parameter. The mem-
bership functions of proposal system are shown in Fig. 3. In Table 1, we show the
Fuzzy Rule Base (FRB) of the proposal system, which consists of 81 rules.
The term sets of CKT, UP, RSSI and NP are defined respectively as:
CKT = {Short, M iddle, Long}
= {S, M, L};
U P = {Low, M iddle, High}
= {Lo, M i, Hi};
RSSI = {Low, M iddle, High}
= {LO, M I, HI}.
N P = {F ew, M iddle, M any}
= {F e, M id, M a};
Effect of Client Priority 239
S M H Lo Mi Hi
1 1
µ(CKT)
µ(UP)
0.5 0.5
0 0
0 900 1800 2700 3600 0 20 40 60 80 100
CKT UP
(a) Connected Keeping Time (b) User Priority
LO MI HI Fe Mid Ma
1 1
µ(RSSI)
µ(NP)
0.5 0.5
0 0
-100 -80 -60 -40 -20 0 0 50 100 150 200 250 300
RSSI NP
(c) Received Signal Strength Indication (d) Number of Packet
0.5
0.0
0 0.2 0.4 0.6 0.8 1
CP
(e) Connection Priority
5 Simulation Results
In this section, we present the simulation results for the proposed system. In our
system, we decided the membership functions by carrying out many simulations.
We show the simulation results in Figs. 4 and 5. We show the relation between
CP and CKT, UP, RSSI and NP. We consider the CKT and NP as a constant
parameters.
In Fig. 4, we consider the CKT value 360. We change the UP value 0 to 100.
In Fig. 4(a), when the UP is increased, the CP is increased. Also, by increasing
RSSI value, the CP is increased. In Fig. 4(b) and (c), we increased the value of
NP to 150 and 270, respectively. We can see that when NP is increased, the CP
is decreased. In Fig. 5, we increase the CKT value to 3240. We can see that by
increasing CKT value, the CP is increased.
We concluded that when CKT, RSSI and UP values increase and NP value
decreases, the clients have high pirority, so they will keep connection to the AP.
CP [unit]
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
UP [unit] UP [unit]
CKT=360 NP=270
1
RSSI=-10
0.9 RSSI=-50
RSSI=-90
0.8
0.7
0.6
CP [unit]
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
UP [unit]
CP [unit]
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
RSSI=-10
0.1 RSSI=-50 0.1
RSSI=-90
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
UP [unit] UP [unit]
CKT=3240 NP=270
1
RSSI=-10
0.9 RSSI=-50
RSSI=-90
0.8
0.7
0.6
CP [unit]
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
UP [unit]
6 Conclusions
In this paper, we presented the effect of client priority in the performance of a
Fuzzy-based WLAN Triage system. We considered four parameters: CKT, UP,
RSSI and NP. We evaluated the systems by simulations. We concluded that
when CKT, RSSI and UP values increase and NP value decreases, the clients
have high pirority, so they will keep connection to the AP. In our future work,
we would like to carry out extensive experiments.
References
1. A particle swarm optimisation algorithm to generate inhomogeneous triangular
cells for allocating base stations in urban and suburban areas
2. Choi, S., Del Prado, J., Mangold, S., et al.: IEEE 802.11e contention-based channel
access (EDCF) performance evaluation. In: International Conference on Commu-
nications (ICC 2003), vol. 2, pp. 1151–1156 (2003)
3. Elmazi, D., Kulla, E., Oda, T., Spaho, E., Sakamoto, S., Barolli, L.: A comparison
study of two fuzzy-based systems for selection of actor node in wireless sensor actor
networks. J. Ambient. Intell. Hum. Comput. 6(5), 635–645 (2015)
4. Elmazi, D., Sakamoto, S., Oda, T., Kulla, E., Spaho, E., Barolli, L.: Two fuzzy-
based systems for selection of actor nodes inwireless sensor and actor networks: a
comparison study considering security parameter effect. Mob. Netw. Appl. 21(1),
53–64 (2016)
5. Gao, D., Cai, J., Ngan, K.N.: Admission control in IEEE 802.11e wireless LANs.
IEEE Netw. 19(4), 6–13 (2005)
6. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A secure-aware call admis-
sion control scheme for wireless cellular networks using fuzzy logic and its perfor-
mance evaluation. J. Mob. Multimed. 11(3&4), 213–222 (2015)
7. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A QoS-aware admission
control system for WLAN using fuzzy logic. In: The 30th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA
2016), pp. 499–505 (2016)
8. Inaba, T., Sakamoto, S., Oda, T., Ikeda, M., Barolli, L.: A testbed for admission
control in WLAN: a fuzzy approach and its performance evaluation. In: The 11th
International Conference on Broadband and Wireless Computing, Communication
and Applications (BWCCA 2016), pp. 559–571 (2016)
9. Javanmardi, S., Shojafar, M., Shariatmadari, S., Ahrabi, S.S.: FRTRUST: a fuzzy
reputation-based model for trust management in semantic P2P grids. Int. J. Grid
Util. Comput. 6(1), 57–66 (2014)
10. Kandel, A.: Fuzzy Expert Systems. CRC Press, Boca Raton (1991)
11. Klir, G.J., Folger, T.A.: Fuzzy Sets, Uncertainty, and Information. Prentice Hall,
Englewood Cliffs (1988)
12. Kolici, V., Inaba, T., Lala, A., Mino, G., Sakamoto, S., Barolli, L.: A fuzzy-based
CAC scheme for cellular networks considering security. In: 2014 17th International
Conference on Network-Based Information Systems, pp. 368–373. IEEE (2014)
13. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improvement
of JXTA-overlay P2P platform: evaluation for medical application and reliability.
Int. J. Distrib. Syst. Technol. (IJDST) 6(2), 45–62 (2015)
Effect of Client Priority 243
14. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improving
reliability of JXTA-overlay P2P platform: a comparison study for two fuzzy-based
systems. J. High Speed Netw. 21(1), 27–42 (2015)
15. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: A comparison
study for two fuzzy-based systems: improving reliability and security of JXTA-
overlay P2P platform. Soft Comput. 20(7), 2677–2687 (2016)
16. Mangold, S., Choi, S., Hiertz, G.R., Klein, O., Walke, B.: Analysis of IEEE 802.11e
for QoS support in wireless LANs. IEEE Wirel. Commun. 10(6), 40–50 (2003)
17. Mangold, S., Choi, S., May, P., Klein, O., Hiertz, G., Stibor, L.: IEEE802.11e
wireless LAN for quality of service. Proc. Eur. Wirel. 2, 32–39 (2002)
18. Matsuo, K., Elmazi, D., Liu, Y., Sakamoto, S., Mino, G., Barolli, L.: FACS-MP: a
fuzzy admission control system with many priorities for wireless cellular networks
and its performance evaluation. J. High Speed Netw. 21(1), 1–14 (2015)
19. McNeill, F.M., Thro, E.: Fuzzy Logic: A Practical Approach. Academic Press,
Boston (1994)
20. Munakata, T., Jani, Y.: Fuzzy systems: an overview. Commun. ACM 37(3), 68–76
(1994)
21. Procyk, T.J., Mamdani, E.H.: A linguistic self-organizing process controller. Auto-
matica 15(1), 15–30 (1979)
22. Qashi, R., Bogdan, M., Hänssgen, K.: Evaluating the QoS of WLANs for the
IEEE802.11 EDCF in real-time applications. In: International Conference on Com-
munications and Information Technology (ICCIT 2011), pp. 32–35 (2011)
23. Romdhani, L., Ni, Q., Turletti, T.: Adaptive EDCF: enhanced service differentia-
tion for IEEE802.11 wireless ad-hoc networks. In: Wireless Communications and
Networking (WCNC 2003), vol. 2, pp. 1373–1378 (2003)
24. Sakamoto, S., Oda, T., Ikeda, M., Barolli, L., Xhafa, F.: Implementation and evalu-
ation of a simulation system based on particle swarm optimisation for node place-
ment problem in wireless mesh networks. Int. J. Commun. Netw. Distrib. Syst.
17(1), 1–13 (2016)
25. Song, N.O., Kwak, B.J., Song, J., Miller, L.E.: Enhancement of IEEE802.11 dis-
tributed coordination function with exponential increase exponential decrease
backoff algorithm. In: The 57th IEEE Semiannual Vehicular Technology Confer-
ence, vol. 4, pp. 2775–2778 (2003)
26. Spaho, E., Sakamoto, S., Barolli, L., Xhafa, F., Ikeda, M.: Trustworthiness in P2P:
performance behaviour of two fuzzy-based systems for JXTA-overlay platform. Soft
Comput. 18(9), 1783–1793 (2014)
27. Wu, H., Peng, Y., Long, K., Cheng, S., Ma, J.: Performance of reliable transport
protocol over IEEE802.11 wireless LAN: analysis and enhancement. In: The 21st
Annual Joint Conference of the IEEE Computer and Communications Societies,
vol. 2, pp. 599–607 (2002)
28. Yang, X., Vaidya, N.H.: Priority scheduling in wireless ad hoc networks. In: Pro-
ceedings of the 3rd ACM International Symposium on Mobile Ad Hoc Networking
& Computing, pp. 71–79 (2002)
29. Zadeh, L.A., Kacprzyk, J.: Fuzzy Logic for the Management of Uncertainty. John
Wiley & Sons, Inc., New York (1992)
30. Zhu, J., Fapojuwo, A.O.: A new call admission control method for providing desired
throughput and delay performance in IEEE802.11e wireless LANs. IEEE Trans.
Wirel. Commun. 6(2), 701–709 (2007)
31. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications. Springer Science &
Business Media, New York (1991)
Energy Saving in HetNet Network
Using eNB Parameters Tuning
1 Introduction
Long Term Evolution (LTE), designed by 3GPP, is enhanced by the use of new
access techniques and by the improvement of architecture that combines differ-
ent new services. New methods are included to LTE to migrate to LTE Advanced
such as Coordinated Multipoint (CoMP), Massive MIMO, carrier aggregation
and HetNet. The notion of LTE Advanced based on HetNet aims to ameliorate
spectral efficiency. HetNet is a combination between various cell types: micro-
cells, picocells and femtocells, and separate access technology [1].
In recent years, data traffic in the wireless mobile network has gained consid-
erable growth. Consequently, energy consumption of Information and Communi-
cation Technology (ICT) is rising with a staggering rate. This significant increase
in ICT energy consumption has many environmental and economic effects. Thus,
on the ecological side, ICT is responsible practically for 2% of CO2 emissions [2].
Moreover, on the economic side, the percentage of ICT energy consumption
varies from 2% to 10% of the annual worldwide energy consumption [3].
Referring to Fig. 1 representing the consumption of power in a typical wireless
mobile network [4], it is clear that the base station (BS) consume the biggest
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 244–255, 2019.
https://doi.org/10.1007/978-3-030-15035-8_23
Energy Saving in HetNet Network Using eNB Parameters Tuning 245
part of the energy that can attain 57 % among all networks [4]. Therefore,
minimize energy consumption in the total network amounts to minimizing energy
consumption in the BS.
Several researches are proposed to reduce energy consumption in cellular
networks, the most well-known are the Switch ON/OFF algorithms. The condi-
tion to apply the Switch ON/OFF procedure is different from one technique to
another. The main aim is to reduce the energy consumption by optimizing the
use of the power of BSs with guaranteed QoS.
In [5], according to the traffic variation, the authors switch ON/OFF dynam-
ically the BSs by respecting certain conditions of blocking probability in order
to guarantee the QoS. In [6], BSs are switched ON/OFF randomly to minimize
energy consumption in a UMTS cellular network. An improvement of this work
is given in [7] where authors proposed a Switch ON/OFF algorithm using a
uniform and a hierarchical scenario.
Algorithms of Switch ON/OFF are also applied in LTE and LTE-A networks
through some researches. Notably, in [8], Alexandra et al. try to come up with
an optimum combination of switching ON/OFF eNBs in order to guarantee the
maximum of energy saving. In [9], Narjes et al. implemented a Switch ON/OFF
scheme in LTE network especially during low traffic period and QoS is guaran-
teed using a power up/down procedure. An improvement of this work is detailed
in [10], where the algorithm of switch ON/OFF is implemented in an LTE-A
based HetNet. In this work, the switch ON/OFF procedure is performed based
on the SINR values of UEs. QoS and coverage are maintained using femtocells
and CoMP techniques.
In all works mentioned above, researchers are concentrated on power con-
sumption issue in UMTS, LTE and LTE-A networks. They all use the Switch
ON/OFF procedure that aims to raise the power saving by minimizing the num-
ber of active eNBs during low traffic periods.
In this paper, we are interested by applying algorithms of Switch ON/OFF in
LTE-A based HetNet. Our contribution is to switch off an eNB when the number
of their attached UEs is lower than a given threshold. UEs of Switched off eNB
are assigned after the Switch OFF depending on their measured SINR (Signal
246 N. Lassoued et al.
to Interference plus Noise Ratio) value. Femtocells will serve UEs with good
SINR value and neighbors eNBs will serve UEs with bad SINR value. Coverage
is maintained by applying a tilt variation procedure for the neighboring eNB.
The rest of the paper is organized as follows: Sect. 2, describes the Net-
work architecture. Section 3, introduces the eNodeB power model. In Sect. 4, we
describe the proposed algorithm. Results of simulation are the subject of Sects. 5
and 6 concludes the paper.
2 Network Architecture
For the network architecture, we propose a HetNet dense urban network, in
which, neighboring cells overlaps. The considered topology is formed by N = 7
eNBs. Each eNB can serve one macrocell surrounded by two femtocells as shown
in Fig. 2. For UEs generation, we apply a non uniform distribution of users where
UEs are generated randomly between cells.
2.1 Femtocells
Due to the use of HetNet, we benefit from the use of femtocells to reduce power
consumption. At first, femtocells are developed to raise capacity and guarantee
coverage for highly densified areas [1]. Moreover, Femtocells are small, inexpen-
sive and are characterized by low energy consumption of BSs [11]. For this reason,
in our work, femtocells are disposed around the eNB to compensate switched off
procedure.
Where, Pa (i) is the eNB i power amplifier including Feeder, Ps (i) is the power
supply for eNB i and Pc (i) is the cooling power consumed by eNB i. The trans-
mitted power PT X (i) is:
Nu
PT X (i) = PT X (j) (3)
j
Where, Nu is the number of eNB i UEs and PT X (j) is the power transmitted
by UE j given by:
PT X (j) = NRB (j).PRB (i) (4)
NRB (j) is the number of resource blocks attributed to UE j and PRB (i) is the
power transmitted per resource block for eNB i. PRB (i) is calculated using the
following equation:
PT X (i)
PRB (i) = (5)
NRB (i)
Where, PT X (i) is the transmitted power defined in Eq. 3 and NRB (i) is the total
number of resource blocks attributed to eNB i. The state of eNB Si (t) is defined
as follows:
1 if ON
Si (t) = (6)
0 if OF F
The total power of the network Ptotal at an instant t is given based on Eq. 1
by:
N
N
Ptotal (t) = [Pi (t)] + ΔP = [POP (i) + PT X (i).Si (t)] + ΔP (7)
i i
Where, Pi is the eNB i power defined in Eq. 1 and ΔP is the amount of added
power after applying the Switch ON/OFF algorithm.
248 N. Lassoued et al.
The received power at UE’s side is calculated using the pathloss model defined
in [13] as:
PRX (θ) = PT X .G(θ).D−α (8)
Where, PT X is the eNB transmission power described in Eq. 4. D is the distance
in km between the eNB and UE, α is the propagation exponent selected depend-
ing on the type of propagation medium and G(θ) is the antenna gain defined
based on equations in [12] by:
arctan (D/h)−θ 2
G(θ) = Gmax .10−12.( θ0 )
(9)
Our energy saving algorithm is based on the concept of Switch ON/OFF espe-
cially during low traffic periods when eNBs are little used. We add to this concept
some new features that significantly reduce energy consumption in the network.
Our algorithm behaves as follow: At first, all macro eNBs are activated, and
all femtocells are deactivated. When the number of active UEs attributed to an
eNB i is smaller than a given threshold T, a Switch OFF procedure is applied to
eNB i, at the same time, femtocells existing in the cell range of switched off eNB
are activated to serve UEs surrounding the deactivated eNB. For UEs located
at the cell extremity, they are allowed to the nearest active eNB neighbor.
Switching ON/OFF eNBs achieve a good amount of power saving but it
affects the QoS by reducing the coverage; Thus, when an eNB is switched off, the
coverage will be decreased and thereby, some UEs will be out of cell’s range. For
this reason, we must compensate the coverage area to provide network coverage
to UEs after switching off process. Thus, to achieve coverage continuity we will
perform a power adjustment procedure combined with a tilt variation technique
of the antenna tilt of neighbor eNBs that will serve the attached UE of the
deactivated eNB. The Switch ON/OFF procedure is stopped when the eNB that
we decide to switch OFF contains some active UEs existing in the cell extremity
and there is no neighbor eNB in ON state. Steps for our proposed approach are:
Noting that Sf (t) is the state of the femtocell at an instant t given by:
1 if ON
Sf (t) = (10)
0 if OF F
• Step 3: After taking the decision of Switch ON/OFF, we should specify new
assignment of UEs of switched off eNB i. The allocation of a UE j is taken
depending on its SINR value. The SINR of a UE j is calculated using the
following formula:
PRX (j)
SIN R(j) = N (11)
i=j PRX (i) + KTp W
N
Where, PRX (j) is the power received at UE j, given in Eq. 8, i=j PRX (i) is
the interference average power. KTp W indicate the background noise, where
K is the Boltzmann constant, Tp is the temperature and W is the bandwidth
used by UE.
The decision to allocate each UE to the correct cell is done referring to
the Table 1.
f = arg
(f )
max (PRX (i, j)) (12)
f ∈Nf emtocells
250 N. Lassoued et al.
Where, Nf emtocells (i) is the number of femtocells existing in the cell range
of eNB i and PRX is the received power at user side j coming from the
femtocell f .
The right eNB neighbor n is chosen according to the following hypothesis:
(n)
= arg
n max (PRX (i, j)) (13)
n∈Nneighbors
Where, Nneighbors (i) is the number of eNB i’s neighbors and PRX is the power
received at user side j coming from eNB neighbor n.
When we activate femtocells and attribute to them UEs of deactivated
eNB, a new amount of power denoted δPf emto is added to the total power
Ptotal . δPf emto is given by:
Nf emto
δPf emto (t) = Sf (t).Pf (14)
f =1
Pf is the power of one femtocell and Sf (t) indicate its state defined in Eq. 10.
The assignment of UEs is described in more details in the following algorithm:
• Step 4: After allocation of UEs to femtocells, UEs in the cell borders are
attributed to their nearest neighbor n . Two mechanisms are performed to
ensure coverage: a power adjustment procedure is applied in parallel with
a tilt variation of the antenna tilt of eNB neighbor n0 . The antenna tilt
variation’s process is performed in order to achieve coverage continuity and
raise cells coverages.
The better value of θ should be chosen to offer coverage for UEs located
in the cell extremity of deactivated eNB. Thus the choice of θ should respect
the following condition:
PRX (θ) < Pmin (15)
If the previous condition is false, no tilt variation and no power addition is
performed.
Energy Saving in HetNet Network Using eNB Parameters Tuning 251
5 Perfermance Evaluation
5.1 Simulation Scenario
To assess performances of our proposed scheme, we use an LTE-Advanced-based
HetNet topology using MATLAB as implementation software. We propose an
urban scenario where our topology is composed by N = 7 macrocells. The dis-
tance between eNBs is D = 500 m, and the eNBs-UEs distance is computed
through their positions in the map. As shown in Fig. 4, in each macro cell we
disposed two femtocells. At instant t = 0, all femtocells are in OFF state.
For the allocation of UEs between cells, we suppose a non-uniform and ran-
dom distribution of UEs between cells where the maximum number of UEs served
by an eNB is fixed to 100 UEs. The variation of the total number of active users
during time is presented in Fig. 5.
For the simulation, we consider the eNB’s maximum transmit power for the
downlink transmission that is recommended by [14]:
252 N. Lassoued et al.
Parameters Values
Bandwidth [MHz] 20
Frequency [MHz] 2140
Propagation model Pathloss model
Distribution of user non uniform
Transmitted power for macro cell [dBm] 35–46
Transmitted power for femto cell [dBm] 5
Distance between macro eNBs [m] 500
Number of resource blocks (RBs) in macro cell 100
Number of resource blocks (RBs) in femto cell 20
represents the energy consumption after the use of our proposed Switch ON/OFF
algorithm.
Looking at the Fig. 6, we notice that in high traffic hours, our proposed
algorithm gives almost the same results as the low traffic algorithm, but in
periods of low traffic, our proposed approach realizes the better power saving
than the case of no switch ON/OFF technique is performed and the cases of low
traffic and random schemes.
The percentage of power saving is the result of Fig. 7. As shown in this figure,
the proposed scheme achieves the maximum amount of power saving than the
random and the low traffic schemes. The percentage of power saving achieved by
our proposed scheme can attain 32% especially during night period from 00:00
am to 05:00 am.
The use of femtocells improves the amount of energy saving; thus, femtocells
are characterized by their low amount of energy consumption. Better yet, the use
of the tilt variation technique increase the amount of power saving after applying
the switch ON/OFF algorithm, thus, it improves the QoS by guarantying an
acceptable coverage to UEs.
254 N. Lassoued et al.
In order to evaluate QoS and coverage, we simulate SINR values before and
after the Switch ON/OFF procedure. Using Eq. 11, we calculate SINR for UEs
existing in the cell edge. The result of SINR calculation is presented in Fig. 8. As
it is shown, the SINR values after the switch ON/OFF procedure increase and
the coverage is guaranteed, Using the concept of femtocells and eNB parameters
tuning let the QoS stable and not lot affected.
Fig. 8. SINR values before and after the Switch ON/OFF procedure
6 Conclusion
The purpose of our work in this paper is to find a better method that can reduce
energy consumption in a LTE-A based HetNet network. Our algorithm is based
on two concept: the first one is the use of femtocells that can help to reduce
the energy consumption in the network thanks to their low energy consumption.
The second concept is the tilt variation technique that ensures an acceptable
coverage and maintains QoS. According to the SINR values before and after
Switch ON/OFF, we can say that our algorithm does not affect the QoS of the
network.
Acknowledgment. This work has received funding from the European Union’s Hori-
zon 2020 research and innovation programme under grant agreement No. 761579 (TER-
APOD).
References
1. http://www.3gpp.org/hetnet. Accessed 15 Nov 2018
2. Global Action Plan Report, An inefficient truth, December 2017
3. http://green-broadband.blogspot.com. Accessed 15 Nov 2018
4. Han, C., et al.: Green radio: radio techniques to enable energy efficient wireless
networks. IEEE Commun. Mag. 49(6), 46–54 (2011)
5. Gong, J., Zhou, S., Niu, Z., Yang, P.: Traffic-aware base station sleeping in dense
cellular networks. In: IWQoS 2018, pp. 1–2, June 2010
Energy Saving in HetNet Network Using eNB Parameters Tuning 255
6. Chiaraviglio, L., Ciullo, D., Meo, M., Marsan, M.A.: Energy-aware UMTS access
networks. In: W-GREEN 2008, Lapland (2008)
7. Chiaraviglio, L., Ciullo, D., Meo, M., Marsan, M.A.: Energy efficient management
of UMTS access networks. In: ITC 21, Paris, France, pp. 1–8, September 2009
8. Bousia, A., Kartsakli, E., Alonso, L., Verikoukis, C.: Energy efficient base station
maximization switch off scheme for LTE-advanced. In: IEEE International Work-
shop CAMAD, Barcelona, Spain, pp. 256–260, September 2012
9. Lassoued, N., Boujnah, N., Bouallegue, R.: Power saving approach in LTE using
switching ON/OFF eNodeB and Power UP/DOWN of Neighbors, SETIT, Tunisia,
December 2018
10. Lassoued, N., Boujnah, N., Bouallegue, R.: Reducing power consumption in HetNet
network using power adjustment and coordinated multipoint technique. In: AINA,
Cracow, Poland, pp. 539–544, May 2018
11. Andrews, J.G., Claussen, H., Dohler, M., Rangan, S., Reed, M.C.: Femtocells: past,
present, and future. IEEE J. Sel. Areas Commun. 30(3), 497–508 (2012)
12. Guidelines for evaluation of radio interface technologies for IMT-Advanced. Radio-
communication Sector of ITU (2009)
13. Salous, S.: Radio Propagation Measurement and Channel Modeling. Wiley,
New York (2013)
14. E-UTRA; further advancements for E-UTRA physical layer aspects, 3GPP (2010)
Robust Road Lane Detection for High
Speed Driving of Autonomous Vehicles
Hyunhee Park(B)
1 Introduction
Various technologies are being researched to improve the safety and convenience
of vehicle driving [1]. In particular, technologies using image information include
pedestrian detection, vehicle detection, obstacle detection, and lane detection
[2,3]. Among these, lane detection techniques have become a core technology
for autonomous vehicles as they can be directly applied to lane departure pre-
vention and lane keeping technologies [4]. Actual lanes are mixtures of straight
and curved lanes which need to be detected quickly. Image-based lane detection
technologies use different algorithms for straight and curved lanes. Therefore,
techniques to quickly predict straight and curved lanes and to determine lanes
are required. The existing methods using straight line models are mainly used
to detect nearby lanes that appear almost straight. Furthermore, methods that
consider curved sections as successions of straight sections divide a lane into
multiple sections by distance; the lanes in the curved sections are then detected
using the fact that each short section would be approximately straight [5,6].
When lanes are assumed to be straight sections, the Hough transform model is
typically used to detect straight lanes. Applying the Hough transform enables
the effective detection of various straight lines in images containing lanes. The
Hough transform method converts points in an image space into a Hough space,
and chooses a section that is highly likely to be a straight line by comparing the
accumulated values. The conventional Hough transform method converts all fea-
ture points in the entire image into a Hough space. When the Hough transform
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 256–265, 2019.
https://doi.org/10.1007/978-3-030-15035-8_24
Robust Road Lane Detection for High Speed Driving 257
is performed in this way, the Hough space creation step takes too much time,
and unnecessary data can be detected. Therefore, to detect lanes quickly during
high speed driving, this study proposes a lane detection method that applies the
vanishing point (VP) update method considering the camera location relative
to the road surfaces and lanes in the Hough space creation process [7,8]. The
proposed method updates the VPs by accumulating only information related to
straight lines passing near the VPs and performs the Hough transform only for
straight lines that pass around the predicted VPs. A model that is appropriate
for high speed driving is then created by predicting the areas in advance for
which the Hough transform will be carried out without measuring all straight
lines and distances.
Figure 1 shows the road image used for lane detection. In general, the horizontal
lines and VPs are located at the center or a little above the center of the image
[9]. A VP here means a point at which the parallel straight lines on the same
plane meet after a projection transformation in the camera. If lane changes in a
general road driving environment are excluded, VPs are located within a certain
area in the images. However, the VPs in a curved section meet at different
points. If a curved line is regarded as a collection of sufficiently short straight
lines, when a lane is divided horizontally into a few sections and then they are
smoothly interconnected, a curved line is created.
Figure 2 shows the first VP result created in a curved section. From the VP
on the triangle formed by the left bottom point, the VP, and the right bottom
point, the height of the triangle was divided by half, and the lane was detected
again. The result is shown in Fig. 3, which is the result of the second VP created
in a curved section. Using the same process, the third VP was created, which is
shown in Fig. 4. When these VPs are processed into Bezier curves and smoothly
interconnected, the VPs for a curve can be formed naturally as the VPs move.
One thing that needs to be considered here is the region of interest (ROI).
For fast image processing and VP prediction, the ROI area is made to decrease
step by step, thereby substantially reducing the computational load. Figure 5
shows the ROI extraction result based on the VPs for the entire image. The
computational load of the operation can be significantly reduced by designating
the red area as the ROI based on the VPs. Figure 6 shows the results of edge
detection by ROI. As a preprocessing step, the noise generated during the edge
extraction was removed with a Gaussian filter (5 * 5) and the image was further
smoothed using a median filter. Then, four points were created by selecting four
coordinates using the ROI, and a polygon was created based on these points. The
computational load was reduced and the accuracy was improved by allowing only
the image inside this polygon to be detected. The processing required for edge
detection can be significantly decreased by reducing the operations used for the
surrounding environment, including lanes in the vicinity. Furthermore, as shown
258 H. Park
in Fig. 7, the second ROI is determined based on the second VP creation area in
which the VP moves. This is reduced to half the size of the first ROI. In other
words, the computational load can be halved compared to the first ROI. Through
the same process, the third ROI in Fig. 8 is determined. Finally, a straight line
is obtained using the Hough transform and the lane is detected and displayed
on the screen.
Robust Road Lane Detection for High Speed Driving 259
acquired from the previous image is used. In this case, the center of the first
image is assumed to be the VP of the previous image. The following equation
used for Hough space generation for the nth image:
γ = (x − xn−1
v ) cos θ + (y − yvn−1 ) sin θ. (1)
Robust Road Lane Detection for High Speed Driving 261
Fig. 8. Example of third ROI extraction result based on the third VP.
4 Experimental Results
To measure the performance of the proposed Hough transform method, the fol-
lowing experiment was performed. The main goal of the proposed method is to
improve the lane detection speed using Hough transform. Therefore, the Hough
transform speed and lane detection speed were measured and compared with
those of existing methods. Furthermore, the lane detection performance is com-
pared with the existing methods to measure the change in accuracy according
to the speed improvement. All the images used in this experiment were taken
262 H. Park
from actual roads and the image size was 960 * 540 pixels. To detect lane fea-
tures from actual road images, a total of 6000 images were obtained including
2000 images each for straight, curved, and lane change sections. For objective
performance measurement, randomly created virtual images were also used and
compared. These images correspond to the preprocessed results of actual road
driving images. Virtual images were used for accuracy measurement and com-
parison. The proportions of straight lines with a distance of 50 pixels or less
from the VP to the straight line were set to 10%, 50%, and 90%. For the slopes
of straight lines, 22 slopes from −4 to 4 were used so that all the components
of the coordinates would be integers. Among these, 10 slopes were randomly
selected and straight lines were drawn. In addition, a total of 6000 images were
created by marking 0, 500, 1000, 1500, and 2000 points at random positions
in each image, similar to the actual situation. Because fewer feature points are
detected in remote distances, zero to 2000 noise elements were added to the
images. In other words, 2000 noise elements were generated for each straight
line. The reason for adding noise is that the results obtained after preprocessing
from the road images are not only lane features. In other words, the amount of
noise can vary by the acquired image and the performance of the preprocessing
module. Therefore, an experimental set of various noise intensities was created
and used to demonstrate that the proposed method is also robust against the
effect of noise regardless of external factors. In Fig. 9, even if the amount of noise
is increased, the processing time of the proposed method increases slowly, while
it increases rapidly in other methods.
The performance of each method was compared based on the Hough space
generation speed measured using actual road images. The average measurement
time was determined using 6000 images for each method. The experimental
results in Fig. 10 show that the proposed method is robust against noise in the
Robust Road Lane Detection for High Speed Driving 263
Fig. 10. Hough space processing time by various road environment (ms).
Hough space generation process and faster than the other methods. The experi-
mental results for the Hough space generation and lane detection process using
6000 images are as follows. The processing time of each method was measured
for 1000 road images of each type, and the average processing times are shown
in Figs. 11 and 12. The computation speed of the proposed method was faster by
at least 2 times and to up to 20 times compared to other methods in all cases.
This result suggests that the proposed method is appropriate for detecting lanes
in road images under high speed driving conditions.
264 H. Park
5 Conclusion
This study proposed a modified Hough transform method and a lane detection
method in a Hough space. The proposed method only generates a Hough space
with straight lines adjacent to the VP of the previous frame. The size of the
Hough space can be reduced because it is generated based on the ROI, thus
significantly reducing the computation time for the Hough transform and lane
detection. The experimental results show that the computation speed of the
proposed method in the Hough transform and lane detection process is approx-
imately twice that of the existing methods. The method proposed in this study
can be useful for lane detection systems installed in vehicles due to its high
computation speed and low memory use.
References
1. Park, H.: Implementation of lane detection for self-driving vehicles using GPS. J.
Korean Inst. Commun. Inf. Sci. 43, 18 (2018)
2. Labayrade, R., Douret, J., Aubert, D.: A multi-model lane detector that handles
road singularities. In: Proceedings of the IEEE Conference on Intelligent Trans-
portation Systems, Toronto, Canada, pp. 1143–1148 (2006)
3. McDonald, J., Shorten, R., Franz, J.: Application of the Hough transform to lane
detection in motorway driving scenarios. In: Proceedings of the Irish Signals and
Systems Conference (2001)
Robust Road Lane Detection for High Speed Driving 265
4. Yadav, S., Patra, S., Arora, C., Banerjee, S.: Deep CNN with color lines model for
unmarked road segmentation. In: Proceedings of the IEEE International Conference
on Image Processing (2017)
5. Guo, C., Mita, S., McAllester, D.: Robust road detection and tracking in challenging
scenarios based on Markov random fields with unsupervised learning. IEEE Trans.
ITS 13, 1338–1354 (2012)
6. Tsai, L., Hsieh, J., Chuang, C., Fan, K.: Lane detection using directional random
walks. In: Proceedings of IEEE Intelligent Vehicles Symposium (2008)
7. Li, Q., Zheng, N., Cheng, H.: Spring robot: a prototype autonomous vehicle and its
algorithms for lane detection. IEEE Trans. Intell. Transp. Syst. 5, 300–308 (2004)
8. Wang, J., Ji, Z., Su, Y.: Unstructured road detection using hybrid features. In:
Proceedings of the International Conference on Machine Learning and Cybernetics
(2009)
9. Yun, S., Guo-Ying, Z., Yong, Y.: A road detection algorithm by boosting using
feature combination. In: Proceedings of the Intelligent Vehicles Symposium (2007)
Study of Beam Power Control
of Ka-Band Multi-beam Broadcasting
Satellite Using Meteorological Data
1 Introduction
Even if we give sufficient rain margin, the most electricity will be wasted in case
of fine weather.
Therefore, we focus on an adaptive control method using multi-beam satellite
[1–3]. In our proposed method, first, the rain effect of the next time on the
satellite link is predicted based on the weather condition in each area. Next,
the transmission power of several beams is controlled based on the prediction.
We have considered the effectiveness of the beam control method under the ideal
condition [4,5]. In the proposed method, it is important to consider the precision
of precipitation data. Therefore, in this paper, we consider the precipitation data
of Rader-AMeDAS (Automated Meteorological Data Acquisition System) for the
prediction.
The structure of this paper is as follows. In Sect. 2, we give the outline of the
proposed method and the satellite model. Based on the explanation, we show the
detail simulation procedure, the method to predict link quality and method to
evaluate the link quality. In Sect. 3, the simulation result in case of using Rader-
AMeDAS data has been shown to consider the effectiveness of the precipitation
data. In Sect. 4, we conclude this paper.
the transmission power of each beam [6]. Therefore, it can be considered that
effective countermeasure against the rain attenuation would be achieved by
changing the transmission power of each beam based on the weather condition
in each area.
Beam 7
Beam 8
Beam 0
Beam 6
Beam 5
Beam 3
Beam 2
Beam 4
Beam 1
1. The precipitation data for the prediction of the link quality of next time
obtained by Rader-AMeDAS is gathered.
2. Based on the data, several beams whose link quality is degraded due to
the rain are selected and the transmission power toward of those beams is
enhanced in addition to the normal power.
3. For the evaluation of the link quality, Rader-AMeDAS measured precipitation
data are used.
Study of Beam Power Control of Ka-Band Multi-beam Satellite 269
In order to predict the rain effect on the link quality, we use following steps.
In this research, we consider the following condition for the evaluation of the
link quality.
270 T. Iwamoto and K. Fujisaki
Current Time
10:00 11:00 12:00
Example 1)
In case of using past 1 hour Measured data
measured precipitation data from 10:00 to 11:00
Example 2)
Predicted data Prediction of link
In case of using after 1 hour
from 11:00 to 12:00 quality of the next time
predicted precipitation data
Example3)
Ideal case of using the time Measured data
precipitation data (ideal case) from 11:00 to 12:00
1. Threshold rainfall
According to the ITU-R recommendations [8], the specific rain attenuation
γ [dB/km] is obtained from;
γ = kRα (2)
Here, R [mm/h] is rain rate. k and α are frequency-dependent coefficients.
In order to evaluate the link quality, we define the threshold rainfall. If the
precipitation in a observation point exceeds the threshold rainfall, the point
is regarded as the incommunicable state. The threshold rainfall shown in
Table 2 is calculated by using Eq. (2) and the parameters shown in Table 1.
The threshold rainfall in case of “Not-enhanced beam” is 7.5 mm/h. On the
other hand, because the electricity is enhanced in addition to the normal
power, the threshold rainfall is 25 mm/h.
2. Beam settings
The diameter of the beam is set to 450 km. The precipitation image data used
for controlling the beam consists of 640 × 640 pixels, and each beam shown in
Fig. 2 has about 20,000 pixels. In the simulation, the pixel in the image data
is assumed as the observation point of the radio wave from the satellite.
3. Evaluation of the link quality
Figure 5 shows the simulation flow for the evaluation of the link quality. In
each observation point (=pixel), based on Eq. (3), the link quality is evaluated
as follows;
rainfall > threshold rainfall : Incommunicable state
(3)
rainfall ≤ threshold rainfall : Communicable state
Study of Beam Power Control of Ka-Band Multi-beam Satellite 271
3 Simulation Results
Precipitation data obtained in 2017 is used for the simulation. Measured precip-
itation data or predicted precipitation data at most 2 h before current time is
used for predicting the link quality.
The availability obtained by using the adaptive control of beams with the pro-
posed technique is shown in Fig. 6. The availability becomes 99.55% in case of
no-countermeasure. From this figure, it has been shown that:
First, we evaluate the annual availability of each beam in case of using measured
precipitation data. The results are shown in Fig. 7. From the results, it has been
shown that:
• the availability of the beams on the south such as beam 0 and beam 1, covering
southern area where it has much rain, gets greatly improved compared with
the no-countermeasure,
• similar to the case of annual availability, the availability degrades as the
precipitation data becomes old.
Next, we evaluate the annual availability of each beam in case of using pre-
dicted precipitation data. The results are shown in Fig. 8. From the results, it
has been shown that:
• the availability tendency is same as case of measured data, though there exists
some variations,
• the degree of degradation of availability in case of using predicted data is
smaller than that of using measured data.
Study of Beam Power Control of Ka-Band Multi-beam Satellite 273
Availability[%]
99.91 99.90184
99.90
99.89
99.88
99.87
99.86
ideal 1 hours ago 2 hours ago
99.80
Availability[%]
99.60
99.40
99.20
99.00
98.80
0 1 2 3 4 5 6 7 8
Beam number
99.80
Availability[%]
99.60
99.40
99.20
99.00
98.80
0 1 2 3 4 5 6 7 8
Beam number
4 Conclusions
In this paper, as a countermeasure against the rain attenuation of the next-
generation broadcasting satellite system using Ka-band, we suggest the technique
using multi-beam satellite. In the proposed method, the multi-beam is controlled
based on the weather condition in each area. We considered the AMeDAS pre-
cipitation image data measured in the past or predicted in the past for the beam
control.
The simulation results show that the availability degrades when using old
data. However, the availability degrades only about 0.1% in the case of using the
past 1 h data. Therefore, it seems that we can control the beams appropriately
by the past 1 h data.
For the future work, we will consider 10 min intervals, such as AMeDAS
10 min precipitation data, in order to realize a situation more similar to the
reality.
References
1. Yoshino, T., Ito, S.: A selection method of beams compensated using AMeDAS data
for the multi-beam satellite broadcasting. IEICE Trans. B J82-B(1), 64–70 (1999).
(in Japanese)
2. Matsudo, T., Karasawa, Y., Shiokawa, T.: The effect of compensation for rain atten-
uation on Earth-satellite links by onboard resource sharing using AMeDAS data.
IEICE Trans. B J76-B2(5), 373–381 (1993). (in Japanese)
3. Fukuchi, H., Sugimoto, T., Tsuzuku, A.: Advanced satellite broadcasting system in
the 21GHz band -Quantitative evaluation of the effect of adaptive satellite power
control. ITE Tech. Rep. 18(45), 13–19 (1994). (in Japanese)
4. Fujisaki, K.: Study on Improvement of Link Quality using AMeDAS precipitation
data for Multibeam Satellite System. “Evaluation of the Down Link Availability
under the Most Suitable Beam Control”. Fukuoka Inst. Technol. Bull. Paper 28,
9–14 (2017). (in Japanese)
5. Iwamoto, T., Fujisaki, K.: Study of the improvement of the link quality using
AMeDAS rainfall data for multibeam satellite broadcasting system. IEICE Techn.
Rep. 118(176), 85–90 (2018). (in Japanese)
6. Kadowaki, N., Suzuki, R.: Overview of the wideband internet engineering test and
demonstration satellite project. J. NICT 54(4), 3–10 (2007)
7. Ozawa, S., Hasegawa, T., Shimada, M., Hirayama, K., Koishi, Y.: Multibeam
antenna system. J. NICT 54(4), 35–43 (2007)
8. Rec. ITU-R P.618-8, Propagation data and prediction methods required for the
design of Earth-space telecommunication systems, ITU (2001)
Numerical Analysis of Optical Duplexer
Composed of Dispersive and Nonlinear
Dielectric in Two-Dimensional Photonic
Crystal Waveguide with Square Lattice
1 Introduction
Photonic crystal (PC) structure is very important for the integration of optical circuit
because of its propagation characteristics based on the photonic band gap (PBG). The
pillar type PC waveguide is basically composed of periodic array of dielectric rods with
a line of defect in free space, as illustrated in Fig. 1. Because of its band gap, the wave
with limited propagation constant can exists in the waveguide. Its propagation char-
acteristics are determined by the refractive index, the lattice period, and optical fre-
quency [1, 2].
Nonlinear optical effect shows many interesting behaviors. Especially, optical Kerr
effect shows self-modulation of the refractive index and self-focusing/defocusing. The
propagation constant is changed by its optical field.
In two dimensional PC structure, the dispersive linear and nonlinear optical
duplexer is composed of waveguide. Use silica of dispersive linear and nonlinear
dielectric material, and consider whether route-switching is possible depending on the
wavelength of input optical signal.
To investigate such PC waveguide, the finite difference time domain (FDTD)
method is used. The FDTD method can be applied to dispersive linear and nonlinear
dielectric material [3].
ð1Þ
ð2Þ
ð3Þ
where E and H are electric and magnetic field and D is the electric displacement. The
space and time discretization are and i; k; n are the number of corresponding
grid. Giving a set of suitable initial conditions and a constitutive equation between
Numerical Analysis of Optical Duplexer Composed of Dispersive 277
D and E to above equations, the latest field is calculated successively as the increase of
time step number.
Consider composite space of free space, linear dielectric material, and dispersive
linear and nonlinear dielectric material.
The electric field in free space is described as follows;
Dny
Eyn ¼ ð4Þ
e0
Dny
Eyn ¼ ð5Þ
e0 er
aL ¼ 2pfL dL ð9Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffi
bL ¼ 2pfL 1 d2L ð10Þ
2pfL ðes e1 Þ
cL ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffi ð11Þ
1 d2L
2pfNL vð3Þ ð1 aÞ
cR ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð14Þ
1 d2NL
278 K. Haari et al.
where e1 and es are relative permittivity of dispersive linear and nonlinear dielectric
material, fL is linear relaxation frequency, dL is linear decaying factor, fNL is nonlinear
ð3Þ
relaxation frequency, dNL is nonlinear decaying factor, v0 is nonlinear susceptibility
and a is weight factor for Kerr effect.
3 Numerical Examples
The distribution ratio of each PORT, the extinction ratio of the dispersive linear and
nonlinear optical duplexer, and the electric field distribution of the photonic crystal
waveguide were obtained. PORT 1 is a PORT to confirm reflection, PORT 2 and
PORT 3 are output PORT. Three types of dispersive linear and nonlinear optical
duplexer structures were used.
The dispersive linear and nonlinear optical duplexer structure is shown in Fig. 1.
Black circle is the silicon of linear dielectric material, blue circle is the silica of
dispersive linear and nonlinear dielectric material. The refractive index of linear
dielectric material n1 ¼ 3:6 as is in Ref. [5]. The refractive index of dispersive linear
and nonlinear dielectric material n2 ¼ 1:5, relative permittivity of dispersive linear and
nonlinear dielectric material e1 ¼ 2:25; es ¼ 5:25, linear relaxation frequency
fL ¼ 63:7 ½THz, linear decaying factor dL ¼ 2:5 104 , nonlinear relaxation fre-
quency fNL ¼ 14:8 ½THz, nonlinear decaying factor dNL ¼ 3:36 101 , nonlinear
ð3Þ
susceptibility v0 ¼ 0:07, and weight factor for Kerr effect a ¼ 0:7 as is in Ref. [3].
The permittivity of free space e ¼ e0 , i.e., no ¼ 1:0. The lattice period a ¼ 551:8 ½nm,
the radius of dielectric material R ¼ 0:2a, the width of waveguide d ¼ 866:8 ½nm are
used. The FDTD discretizations are Dx ¼ Dz ; 9:85 ½nm and Dt ; 0:0209 ½fs,
respectively. Analysis area is 10:244 ½lm 14:1052 ½lm. As an absorbing boundary,
Berenger’s perfectly matched layer is installed.
The input optical signal is Gaussian beam and given by
( )
x 2 2pct
Ey ðx; tÞ ¼ E0 exp sin ð15Þ
w0 k
Ey ðx; tÞ
Hx ðx; tÞ ¼ qffiffiffiffi ð16Þ
e0 le00
where the amplitude E0 ¼ 1:0 ½V=m, the width of beam spot w0 ¼ 275:8 ½nm, speed
of optical in vacuum c ¼ 2:998 108 ½m=s, the wavelength k ¼ 1:350 1:600 ½lm,
the permittivity of free space e0 ¼ 8:854 1012 ½F=m, and the magnetic permittivity
of free space l0 ¼ 4p 1017 ½H=m.
The distribution ratio of each PORT is described as follows;
PPORTi
Si ¼ 100 ði ¼ 1; 2; 3Þ ð17Þ
PPORT1 þ PPORT2 þ PPORT3
The extinction ratio of the dispersive linear and nonlinear optical duplexer is
described as follows;
S2
10 log10 ð18Þ
S3
100.000
90.000
80.000
Distribution ratio [%]
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
10.000
8.000
6.000
Extinction ratio [dB]
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
Fig. 3. Extinction ratio of dispersive linear and nonlinear optical duplexer in Fig. 1.
(a) (b)
Fig. 4. Electric field distribution at (a) k ¼ 1:390 ½lm and (b) k ¼ 1:553 ½lm. Coordinate axis
is in Fig. 1.
Numerical Analysis of Optical Duplexer Composed of Dispersive 281
100.000
90.000
80.000
Distribution ratio [%]
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
Figure 6 shows the distribution ratio, Fig. 7 shows the extinction ratio.
10.000
8.000
6.000
Extinction ratio [dB]
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
Fig. 7. Extinction ratio of dispersive linear and nonlinear optical duplexer in Fig. 5.
(a) (b)
Fig. 8. Electric field distribution at (a) k ¼ 1:515 ½lm and (b) k ¼ 1:434 ½lm.
100.000
90.000
80.000
Distribution ratio [%]
70.000
60.000
50.000
40.000
30.000
20.000
10.000
0.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
14.000
12.000
10.000
8.000
Extinction ratio [dB]
6.000
4.000
2.000
0.000
-2.000
-4.000
-6.000
-8.000
-10.000
-12.000
-14.000
1.350E-06 1.400E-06 1.450E-06 1.500E-06 1.550E-06 1.600E-06
Wavelength [m]
Fig. 11. Extinction ratio of dispersive linear and nonlinear optical duplexer in Fig. 9.
(a) (b)
Fig. 12. Electric field distribution at (a) k ¼ 1:451 ½lm and (b) k ¼ 1:561 ½lm.
Numerical Analysis of Optical Duplexer Composed of Dispersive 285
4 Concluding Remarks
According to the results of this research, it was confirmed that dispersive linear and
nonlinear optical duplexer can switch the route according to the wavelength of the
input optical signal. By changing the structure of dispersive linear and nonlinear optical
duplexer, the distribution ratio, the extinction ratio, and the wavelength difference of
switching the route changed.
It is necessary to narrow the wavelength difference of switching the route, and
increase the numerical value of the extinction ratio.
References
1. Joannopoulos, J.D., Meade, R.D., Winn, J.N.: Photonic Crystals. Princeton University Press,
New Jersey (1995)
2. Noda, S., Baba, T. (eds.): Roadmap on Photonic Crystals. Kluwer Acedemic Publishers, The
Netherlands (2003)
3. Sullivan, D.M.: Nonlinear FDTD formulations using Z transforms. IEEE Trans. Microw.
Theory Tech. 43(3), 676–682 (1995)
4. Yee, K.S.: Numerical solution of initial boundary value problems involving Maxwell’s
equation. IEEE Trans. Antennas Propag. 14(3), 302–307 (1996). IEEE Trans. EMC 32(3),
222–227 (1990)
5. Li, X., Maeda, H.: Numerical analysis of photonic crystal directional coupler with kerr-type
nonlinearity. In: Proceedings of 5th Asia-Pacific Engineering Research Forum on Microwaves
and Electromagnetic Theory, pp. 165–168 (2004)
Cycle Interference Alignment for the Full
Duplex Communication System Based
on User Virtual Splitting Thought
1 Introduction
The mobile communications for the future has become a global hot spot of research and
development. The efficient use of spectrum resources is the main goal in the design of
communication system. The new duplex technology provides the possibility to achieve
this goal. Particularly, the full duplex (FD) technology, which refers to the simulta-
neous operation of the transmitter and receiver at the same frequency, is an accepted
way to improve the spectrum efficiency of the system. In theory, the spectral efficiency
of the FD wireless communication systems can be doubled [1]. However, when the
technology of the FD is used in the communication system, there is a lot of the
additional interference. For example, the uplink users will produce interference in the
downlink users. If this interference is not canceled, the FD system cannot obtain two
times the spectral efficiency of the half duplex (HD) system [2].
At present, the research results for single-cell single-user full-duplex system are
relatively perfect [3]. However, for the single-cell multi-user full-duplex system, the
relative research is not quite comprehensive [4]. In the paper [4], it supposes that the
transmitting user does not interfere with the receiving user. Obviously, the result is too
optimistic. In addition, the complexity of the scheme is very high, especially when the
number of antennas in the node is relatively large. Therefore, the present algorithm is
not suitable for the actual system.
In recent years, the scholars have found other interference management mecha-
nisms besides interference elimination and interference orthogonalization. The most
representative of them is the interference alignment (IA) [5, 6]. The basic idea of IA is
to design the space of sending signals, and all the interference signals of each receiver
are compressed into the smallest subspace. In this way, the larger signal space is
reserved to receive the target signal [7]. The degree of freedom (DoF) is the first
approximation of system data rate [8]. It is an important index to measure IA tech-
nology. Its physical meaning is the number of data flows that the system can transmit at
the same time. From the perspective of DoF, we can see that IA can make the capacity
of the system increase linearly with the number of users. It proves that the capacity of
multi-user system is not limited by interference. This has broken the past cognition that
the wireless communication system is the interference limited system. Therefore, IA
has aroused extensive attention and research in the academic circle.
In this paper, the IA technology is explored to solve the interference from the intra-
user and the inter-users in the single-cell FD-BS-FD-user system. Firstly, the user
virtual splitting thought is introduced, and the system is equivalent to the FD-BS-HD-
user system. Then, a set of cycle IA equations are designed to reach the optimal system
degree of freedom (DoF). The simulation shows that the cycle IA scheme can reach the
desired result.
The system model of the single-cell multi-user full-duplex system is shown as Fig. 1.
The system parameters configured is (M, N, K). Both the base station (BS) and users
work in duplex mode. There are M antennas in the BS and N antennas in each user.
There are K users. Obviously, for each users, there are interference from inter-users, as
well as interference from the uplink signal of the intra-user. Gi,j is the interference
channel from user j to user i.
Since the base station cannot support M or more users in the same time, the
parameters in this system are supposed as M > K. In addition, we suppose the
parameters of (M, N, K) can satisfy the feasibility condition of IA [9].
For the FD-BS-FD-user system, to deal with interference problems, we adopt the cyclic
interference alignment algorithm based on the user virtual splitting thought. Firstly,
each full-duplex user is equivalently split into one virtual uplink user and one virtual
downlink user. Then, the cyclic interference alignment algorithm is adopted to elimi-
nate the interference from the intra-user and the inter-users. Finally, the whole system
reach the maximum DoF, that is, the maximum system capacity.
H i[ u ] H i[ d ] H j[ d ] H j[u ]
(i, j = 1, , K)
G i, j
G j ,i
BS-HD-user system where K[u] = K[d] = K. In this way, for each virtual downlink user,
the inference signals come from all the K virtual uplink users.
uplink signal
interference tx antennas rx antennas
1 N 1 N
tx antennas rx antennas .. .. tx antennas rx antennas
1 N 1 N 1 N 1 .. N
.. .. ..
...
UE 2
UE 1 UE K
inter-user interference signal
1 N 1 N
.. ..
..
..
virtual uplink virtual downlink
user 2 user 1
...
...
.. ..
virtual uplink virtual downlink
user 3 user 1
½u ½u
G1;1 V 1 ¼ t1 G1;2 V 2
½u ½u
G2;2 V 2 ¼ t2 G2;3 V 3
...
½u ½u
GK1;K1 V K ¼ tK1 GK1 ;K V K
½u ½u
GK;K V K ¼ tK GK;1 V 1
½u ½u
For each downlink user j, the interference space is spanðGj;1 V 1 ; Gj;2 V 2 ; . . .;
½u
Gj;K V K Þ.According to the feasibility condition of IA, when K[u] = K[d], the minimal
configuration of N should satisfy Nmin = max{K[u], 2}. Therefore, we can get
½u ½u ½u
rankðGj;1 V 1 ; Gj;2 V 2 ; . . .; Gj;k V K Þ ¼ K ½d 1\N. In this way, every downlink user
can receive one DoF at least without interference.
290 H. Ren et al.
4 Simulation Results
In Fig. 3, we show the simulation results of the sum rate of the FD-BS-FD-user system
in different system configuration. It can be seen that the sum rate increases with the
number of the users. In other words, the cycle IA scheme can reach the desired results.
70 K=5
K=4
60 K=3
Average Sum Rate [bps/Hz]
50
40
30
20
10
0
0 5 10 15 20 25
SNR [dB]
Fig. 3. Sum rate versus different system configurations
70 Cycle IA
ZF-MRC
60
Average Sum Rate [bps/Hz]
50
40
30
20
10
0
0 5 10 15 20 25
SNR [dB]
Fig. 4. Sum rate versus different algorithms
5 Conclusions
In this paper, we focus on the single-cell multi-user full-duplex system. By using the
cycle IA, the sum rate of the system increases with the system parameter configuration.
The method in this paper can be promoted to the case that the number of the user data
rate more than one. Furthermore, it can reach the maximum sum-DoF, and satisfy the
transmission requirements of system uplink and downlink.
Acknowledgments. This work was supported by the Beijing Natural Science Foundation
(4172071), Youth Project of China Academy of Information and Communications Technology.
References
1. Sabharwal, A., Schniter, P., Guo, D., Bliss, D.W., Rangarajan, S., Wichman, R.: In-band full-
duplex wireless: challenges and opportunities. IEEE J. Sel. Areas Commun. 32(9), 1637–
1652 (2014)
2. Goyal, S., Liu, P., Panwar, S., DiFazio, R.A., Yang, R., Bala, E.: Full duplex cellular systems:
Will doubling interference prevent doubling capacity? IEEE Commun. Mag. 53(5), 121–127
(2015)
292 H. Ren et al.
3. Nguyen, D., Tran, L.-N., Pirinen, P., Latva-aho, M.: Precoding for full-duplex multiuser
MIMO systems: Spectral and energy efficiency Maximization. IEEE Trans. Signal Process. 61
(16), 4038–4050 (2013)
4. Nguyen, D., Tran, L.-N., Pirinen, P., Latva-aho, M.: On the spectral efficiency of full-duplex
small cell wireless systems. IEEE Trans. Commun. 13(9), 4896–4910 (2014)
5. Cadambe, V.R., Jafar, S.A.: Interference alignment and degrees of freedom of the, K-user
interference channel. IEEE Trans. Inf. Theory 54(8), 3425–3441 (2007)
6. Kim, D., Torlak, M.: Timization of interference alignment beamforming vectors. IEEE J. Sel.
Areas Commun. 28(9), 1425–1434 (2010)
7. Sahai, A., Diggavi, S., Sabharwal, A.: On degrees-of-freedom of full-duplex uplink/downlink
channel. In: Proceedings of the IEEE Information Theory Workshop, Seville, Spain (2013)
8. Jafar, S.A., Fakhereddin, M.J.: Degrees of freedom for the MIMO interference channel. IEEE
Trans. Inf. Theory 53(7), 2637–2642 (2007)
9. Kim, K., Jeon, S.W., Kim, D.K.: The feasibility of interference alignment for full-duplex
MIMO cellular networks. IEEE Commun. Lett. 19(9), 1500–1503 (2015)
The 12th International Symposium on
Mining and Web (MAW-2019)
Proposal of Web API Architecture
for Smart Community: A Case Study
of Japan
Kanae Matsui(B)
1 Introduction
A smart community consists of numerous Internet of Things (IoT) devices and
handles massive IoT data for providing data-driven services to residents who are
owners of the data. Setting up IoT devices in a community involves high cost, and
making benefits from IoT data, which are collected by IoT devices, is necessary
to manage a smart-community-based IoT platform. Therefore, the IoT platform
should have a Web API (Application Programming Interface) to provide IoT
data to a third party that provides IoT-based services and applications. However,
only providing IoT data is not enough to enhance their business, because the
extraction of meaningful information from IoT data requires an understanding
of the background.
From this context, this study proposes a Web API system and utilizes a col-
lected by IoT devices that are set in Urawamisono, Saitama city, Japan, where is
proceeding to be smart community [1]. In addition, the system puts background
information, which needs to understand data meaning, into the data. In this
time, indoor temperature is utilized as a sample to show a way how to utilize
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 295–304, 2019.
https://doi.org/10.1007/978-3-030-15035-8_28
296 K. Matsui
2 Related Work
Two software, including a Web API that handles IoT for data-driven systems,
which consist of a smart city as their component, are introduced. Our target is
a “smart community,” which is smaller than a city, but the target is the same in
terms of IoT data utilization. In this section, representative Web API functions
and differences from the proposed Web API.
2020 Summer Olympic Games, prompting local and regional officials to consider
how they will accommodate the massive influx of event spectators and maximize
the impact of this influx for broader development goals.
This paper proposes a Web API for this developing smart community, which
has the following differences from above-mentioned smart-city APIs. The above-
mentioned APIs were designed for targeted services, e.g., sightseeing navigation,
traffic congestion avoidance, and housing services. We also aim to distribute IoT
data for third parties that would like to conduct such services by using IoT data.
Interrupting IoT data requires expert knowledge for each data, but it takes a
cost. Therefore, it is necessary to add interrupting data into distributed IoT
data, e.g., temperature prediction. Therefore, the proposed Web API distributes
interruption information into IoT data to enhance third party developing smart
community services.
3 System Description
In this section, a web API system for IoT data utilization based on a smart
community is described. Based on the SCIP project, the SCIP database, which
stores all IoT data in the community, has an interface to access the data from
the third party. The SCIP database has collected various types of IoT data, e.g.,
indoor temperature, humidity, CO2 emission, luminance, PM2.5, etc. The indoor
temperature and humidity can be used for many services, such as smart house
functions, which maintain indoor comfort with an automatic HVAC (Heating,
Ventilation, and Air Conditioning) system. Therefore, the proposed API targets
both indoor temperature and humidity data in this case. Figure 1 illustrates a
conceptual diagram of the Web API.
This Web API has a function to add expert knowledge such as appropriate
value and influence on human body as interpretation information when utiliz-
ing IoT data gathered in the smart community and making it into service. In
this research, in order to deal with This Web API has a function to add expert
298 K. Matsui
Request Parameter
All temperature data of /simple?sensor=temp+home id=1
Building ID 1
From 2017−2−25 to current /custom?sensor=temp+home id=1
Building ID 1
All temperature data up to start=2017−2−25+end=2017−10−3
2017−10−3
Temperature data acquired /now?sensor=temp+home id=1
at the time closest to the
current time of building ID 1
The proposed Web API has the function of attaching expert knowledge, such
as appropriate value and influence on human body, as interpretation information
when utilizing IoT data gathered in the smart community and turning it into
300 K. Matsui
a service by the third party as mentioned previously. In order to deal with the
indoor temperature data, the system investigates the appropriate temperature
for each season and creates a conditional branch. An interpretation sentence is
determined by this conditional branch, and this is combined with the IoT data.
The appropriate temperature for each season is given in Table 4. The conditional
branch is defined by the Japanese season.
In order to distribute IoT data and interpretation data as digital data, the
distribution data is defined in the JSON format. The response fields are given
in Table 5.
4 Evaluation
In order to verify the system, dummy data of indoor temperature was registered
in the constructed database. For confirmation, to start the browser on the client
server connected to the same network as the application server, and input the
URI according to the purpose in the address bar. If the response is correct,
Proposal of Web API Architecture for Smart Community 301
displaying the data in the JSON format would be on the browser of the client
server.
In a Web API, a response time for the third party is necessary to provide IoT
data. In order to verify the response time of this Web API, Postman (ver.5.5.0)
is used [6]. Postman is a Web API testing tool and used for measurement of the
response time. Table 6 lists the information of equipment used for the evaluation.
As a test, three acquisition methods are executed 30 times by the client server
and the result is verified.
When acquiring data by each method, it shows the URI entered in the address
bar of the browser on the client server and displays the result in a web browser.
5 Results
5.1 Item 1
The item 1 is to check when user sent a request to proposed web API with spec-
ifying the building ID and sensor type to acquire all data. Following URI is a
request to acquire all data. The URI includes that IP address, sensor type, and
building ID.
http://192.168.2.45:4567/simple?sensor=temp&home/ id=1
Accordingly, all data-acquisition methods of this Web API were verified and
it was confirmed that the results were displayed correctly. The result of JSON
files in a web browser is shown in Fig. 3.
302 K. Matsui
5.2 Item 2
The item 2 is to check when user sent a request to proposed Web API with
specifying the ID of the building and sensor type. The URI for data request
consists of IP address, sensor type, building ID, which means home ID in this
case, and data when users want to acquire. Figure 4 shows the result of this
request, and then the result showed all requested data were acquired from the
proposed Web API by this type of requests.
5.3 Item 3
The item 3 is to check when user sent a request to proposed Web API with
specifying the ID of the building and the sensor type and acquiring the data
closest to the current time. Following URI is a request to acquire the most
current data.
Figure 5 shows the result of this request, and then the result showed all
requested data were acquired from the proposed Web API by this type of
requests.
In conclusion, the proposed Web API system results showed that all requests
to acquire IoT and interpretation data were completed. However, the response
time took a time more than a request from some services and applications. As
the next step, the response time of the system should be improved to accept all
requests from databases of third parties.
6 Conclusion
In this research, we proposed and implemented the Web API as a data access
interface for creating services and applications using IoT data accumulated in a
smart community. We constructed a database and Web API that are supposed to
be used in the smart community and evaluated this Web API in terms of validity
and response time. Accordingly, in validation verification, we were able to receive
the correct response to the request according to the purpose. In response-time
measurement, it was a Web API with high immediate responsiveness.
From this result, it was thought that it takes more time to acquire data when
operating in the actual smart community. Therefore, in order to use enormous
IoT data by using the Web API, there is a need to design with consideration of
functionality and response time. In this research, the design was realized with
the method of acquiring all specified sensor data, but by making it possible to
set the upper limit of data acquisition number, it is considered that the response
time could be shortened without impairing the functionality.
Acknowledgements. This work was supported in part by the R&D project “Design
of Information and Communication Platform for Future Smart Community Services”
by the Ministry of Internal Affairs and Communications (MIC) of Japan, and was sup-
ported by Funds for the Integrated Promotion of Social System Reform and Research
and Development, MEXT, Japan, and by MEXT/JSPS KAKENHI Grant (B) Number
17K12789 and SECOM Science and Technology Foundation.
References
1. Miura, T., Wijekoon, J.L., Prageeth, S., Nishi, H.: Novel infrastructure with common
API using docker for scaling the degree of platforms for smart community services.
In: 2017 IEEE 15th International Conference on Industrial Informatics (INDIN),
pp. 474–479. IEEE (2017)
2. Kanae, M.: Pleasant sleep provision system during warm nights as a function of
smart home. In: IEEE 2017 International Symposium on Networks, Computers and
Communications (ISNCC), pp. 1–6. IEEE (2017)
304 K. Matsui
3. Erno, K., Martin, B., Jaeho, K., et al.: Standards-based worldwide semantic inter-
operability for IoT. IEEE Commun. Mag. 54(12), 40–46 (2016)
4. Ernoe, K., Martin, B.: Apparatus and method for connecting at least two systems
by converting data. Google Patents, US Patent App. 15/742,069 (2018)
5. General Electric Company, CityTM IoT Platform. https://developer.currentbyge.
com/cityiq. Accessed 25 Jan 2019
6. Postman, Inc. Postman. https://www.getpostman.com/. Accessed 25 Jan 2019
Evaluation Measures for Extended
Association Rules Based on Distributed
Representations
Tomonobu Ozaki(B)
1 Introduction
Frequent itemset mining [1,2], i.e., the enumeration problem of all itemsets
appearing frequently in databases is one of the most fundamental problems in
data mining. As similar, discovery of association rules that consist of two fre-
quent itemsets is the most traditional method in this field. Various extensions
of frequent itemsets and association rules in a wide variety of application fields
have been developed over a long period. Since an association rule consists of
two frequent itemsets, association rule discovery has the same drawback as the
frequent itemset mining has. That is combinatorial explosion. A huge number
of itemsets and rules will be derived if the target database to be mined is dense
or the minimum threshold is set low. To suppress insignificant results, several
techniques have been proposed such as mining condensed representations [3,4]
and filtering based on statistical tests [5,6]. In addition, to rank the results based
on certain viewpoints is also useful in practice, a number of evaluation measures
have been proposed [7,8] including Lift [9] and Conviction [10]. Especially in
recent years, various evaluation measures based on distributed representation
[11] have been proposed. For example, some methods are proposed [12,13] to
evaluate the degree of representatively, exceptionality and significance of fre-
quent itemsets based on the distances among frequent itemsets in a vector space
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 305–313, 2019.
https://doi.org/10.1007/978-3-030-15035-8_29
306 T. Ozaki
2 Preliminary
Let I = {i1 , i2 , · · · , i|I| } be a set of all items. A transaction t ⊆ I is a set of
items in I, and a database D = {t1 , t2 , · · · , t|D| } is a set of transactions. A
pattern P ⊆ I is a combination of items which appears in plural transactions in
common. The support value of P with respect to D is formally defined as
supD (P ) = | {t ∈ D | t ⊆ P } | / |D| .
of frequent patterns in D.
For two frequent patterns A and C having no item in common (A ∩ C = ∅),
we consider a rule in the form of A → C representing that if A holds then C also
holds. The support and confidence values of the rule A → C are respectively
defined as
Given an action rule r1 ⇒ r2 , the support and confidence values are respec-
tively defined as follows [16].
Given the minimum support threshold σ (0 < σ ≤ 1) and the minimum con-
fidence threshold θ (0 < θ ≤ 1), if an action rule r1 ⇒ r2 satisfies the conditions
supD (r1 ⇒ r2 ) ≥ σ and confD (r1 ⇒ r2 ) ≥ θ, then it is called an association
action rule.
−−−−−→ −−−−−→
EIgain (M, x, y) = dist ( −
→
x, →
−
y ) − dist M ∪ {x}, M ∪ {y}
−−−−−→
−−−−−→
EIavg (M, x, y) = 1
2 dist →−
x , M ∪ {y} + dist −→
y , M ∪ {x}
→
−
→
−
EIavg (M, x, y) = 1
2 min dist →
−
x, i + min dist →
−
y, i
i∈M ∪{y} i∈M ∪{x}
The measure EIgain represents the difference in distance between the case of
considering the condition M and that of not considering M . It gives high value
for rules if the distance between x and y becomes smaller in the consideration of
M . The idea behind EIavg is that the indirect association rules are worthwhile
if it is difficult to guess the consequences of one rule from the other rule. By
regarding the distance as the degree of difficulty for the estimation, the measure
EIavg employs the average value of difficulties between the case for estimating x
from M ∪ {y} and that for estimating y from M ∪ {x}. The measure EIavg is a
variant of EIavg , and it measures the difficulty for estimating the consequences
by using each item in the rules separately.
Both measures are based on the difference in distance between the change in
consequences and that in antecedents. They give a high value to an association
action rule whose change in consequences is large but that in antecedents is
small. While EAg evaluates the rules from more global viewpoint by captures
the changes at the level of itemsets, EAl determines the evaluation locally by
considering each attribute separately.
4 Experiments
4.1 Experiments for Indirect Association Rules
To assess the proposed evaluation measures for indirect association rules, we con-
ducted preliminary experiments using Niconico dataset1 provided by DWANGO
Co., Ltd. and National Institute of Informatics in Japan. It is a real dataset on
Japanese video-sharing site Nicovideo2 . In the experiments, each set of tags asso-
ciated with each movie posted in Nicovideo is employed as a transaction. From
the entire dataset, a database Dmusic having about 456 thousands of transac-
tions was prepared by collecting transactions having a tag “music”. Using the
confidence value as a dependency measure, we obtained 1,162 of indirect associ-
ation rules under the settings of σmin = 0.0001, σmax = 0.00001, and τ = 0.15.
As a distributed representation of items, we derive d-dimensional real vectors
where d ∈ {50, 150} using GloVe [19]3 .
To assess the proposed three measures EIgain , EIavg and EIavg , we cal-
culated the basic statistics of values of each evaluation measure for the set of
indirect association rules. The results are summarized in Table 1.
Table 1. The basic statistics of evaluation measures for indirect association rules
d = 50 d = 150
EIgain EIavg EIavg EIgain EIavg EIavg
Minimum value 0.050 0.089 0.095 0.067 0.086 0.104
First quartile 0.205 0.227 0.214 0.218 0.283 0.260
Median 0.261 0.277 0.254 0.286 0.323 0.305
Third quartile 0.327 0.316 0.298 0.343 0.359 0.342
Maximum value 0.489 0.486 0.470 0.489 0.481 0.473
Average 0.263 0.273 0.255 0.279 0.312 0.298
Standard deviation 0.087 0.071 0.063 0.083 0.070 0.068
From the results, it is observed that the measure EIgain has wider range and
larger standard deviation compared with other measures. Although the value of
1
https://www.nii.ac.jp/dsc/idr/nico/nico.html.
2
http://www.nicovideo.jp/.
3
https://nlp.stanford.edu/projects/glove/.
Evaluation Measures for Extended Association Rules 311
EIgain can become negative in theory, the minimum value exceeded zero in the
experiments. We believe that this results come from that the conditions M are
sufficiently considered at the stage of extracting indirect association rules.
d = 50 d = 150
EIgain – EIavg 0.500 0.402
EIgain – EIavg −0.033 0.076
EIavg – EIavg 0.684 0.775
The Nursery Data Set4 was used in the experiments. This data contains 12,960
transactions with nine attributes. While three attributes are treated as stable
attributes, the remained six attributes are used as flexible. After deriving asso-
ciation action rules using σ = 0.0005 and θ = 0.70, we selected 41,727 rules
which have up to three flexible attributes in the antecedent, at least one stable
attribute in the antecedent and one flexible attribute in the consequence. The
distributed representation of items were derived by the same method as in the
case of the experiments for indirect association rules.
Table 3. The basic statistics of evaluation measures for association action rules
d = 50 d = 150
EAg EAl EAg EAl
Minimum value −0.624 −0.912 −0.661 −0.924
First quartile −0.170 −0.154 −0.187 −0.150
Median 0.037 0.098 0.043 0.114
Third quartile 0.121 0.141 0.132 0.157
Maximum value 0.433 0.439 0.439 0.444
Average −0.026 −0.001 −0.026 0.011
Standard deviation 0.180 0.204 0.191 0.208
4
https://archive.ics.uci.edu/ml/datasets/nursery.
312 T. Ozaki
We calculated the basic statistics of values of EAg and EAl for the set of
association action rules. The results are summarized in Table 3.
In the experiments, from the influence of the constraint on the form of associ-
ation action rules, most of rules had negative values irrespective of the evaluation
measures. It is recognized that the measure EAl has wider range and larger stan-
dard deviation. Although not shown in detail, a number of outliers were detected
in EAl . The correlation between EAg and EAl is very high. It is 0.908 for d = 50
and 0.918 for d = 150. In addition, the top ranked association action rules with
respect to the each evaluation measure were the same. From the results, no large
difference between two measures was observed unfortunately.
5 Conclusion
Acknowledgements. In this paper, the author used the “Nicovideo dataset” pro-
vided by National Institute of Informatics. This work was partially supported by JSPS
KAKENHI Grant Number 17K00315.
References
1. Agrawal, R., Imielinski, T., Swami, A.: Mining association rules between sets of
items in large databases. In: Proceedings of the 1993 ACM-SIGMOD International
Conference on Management of Data, pp. 207–216 (1993)
2. Han, J., Cheng, H., Xin, D., Yan, X.: Frequent pattern mining: current status and
future directions. Data Min. Knowl. Disc. 15(1), 55–86 (2007)
3. Boulicaut, J.-F., Bykowski, A., Rigotti, C.: Free-sets: a condensed representation of
boolean data for the approximation of frequency queries. Data Min. Knowl. Disc.
7(1), 5–22 (2003)
4. Pasquier, N., Bastide, Y., Taouil, R., Lakhal, L.: Discovering frequent closed item-
sets for association rules. In: Proceedings of the 7th International Conference on
Database Theory, pp. 398–416 (1999)
5. Webb, G.I.: Discovering significant patterns. Mach. Learn. 68(1), 1–33 (2007)
6. Hämäläinen, W., Webb, G.I.: Statistically sound pattern discovery. In: Proceedings
of the 20th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, p. 1976 (2014)
7. Wu, T., Chen, Y., Han, J.: Re-examination of interestingness measures in pattern
mining: a unified framework. Data Min. Knowl. Disc. 21(3), 371–397 (2010)
Evaluation Measures for Extended Association Rules 313
8. Lenca, P., Vaillant, B., Meyer, P., Lallich, S.: Association rule interestingness mea-
sures: experimental and theoretical studies. In: Guillet, F.J., Hamilton, H.J. (eds.)
Quality Measures in Data Mining, pp. 51–76 (2007)
9. Brin, S., Motwani, R., Silverstein, C.: Beyond market baskets: generalizing asso-
ciation rules to correlations. In: Proceedings of the 1997 ACM SIGMOD/PODS
Joint Conference, pp. 265–276 (1997)
10. Brin, S., Motwani, R., Ullman, J.D., Tsur, S.: Dynamic itemset counting and impli-
cation rules for market basket data. In: Proceedings of the 1997 ACM SIGMOD
International Conference on Management of Data, pp. 255–264 (1997)
11. Bengio, Y., Courville, A., Vincent, P.: Representation learning: a review and new
perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35(8), 1798–1828 (2013)
12. Kawanobe, S., Ozaki, T.: Extraction of characteristic frequent visual patterns by
distributed representation. In: Proceedings of the 2017 31st International Con-
ference on Advanced Information Networking and Applications Workshops, pp.
525–530 (2017)
13. Kawanobe, S., Ozaki, T.: Experimental study of characterizing frequent itemsets
using representation learning. In: Proceedings of the 2018 32nd International Con-
ference on Advanced Information Networking and Applications Workshops, pp.
170–174 (2018)
14. Ozaki, T.: Evaluation measures for frequent itemsets based on distributed represen-
tations. In: Proceedings of the 2018 Sixth International Symposium on Computing
and Networking, pp. 153–159 (2018)
15. Tan, P.-N., Kumar, V., Srivastava, J.: Indirect association: mining higher order
dependencies in data. In: Proceedings of the 4th European Conference on Principles
of Data Mining and Knowledge Discovery, pp. 632–637 (2000)
16. Ras, Z.W., Dardzinska, A., Tsay, L.-S., Wasyluk, H.: Association action rules. In:
Proceedings of the 2018 IEEE International Conference on Data Mining Workshop,
pp. 283–290 (2008)
17. Kawaguchi, M., Ozaki, T.: Finding replaceable ingredients by indirect association
rules. In: Proceedings of the 78th National Convention of Information Processing
Society of Japan, vol. 1, pp. 525–526 (2016). (in Japanese)
18. Suzuki, E.: Discovering action rules that are highly achievable from massive data.
In: Proceedings of the 13th Pacific-Asia Conference on Advances in Knowledge
Discovery and Data Mining, pp. 713–722 (2009)
19. Pennington, J., Socher, R., Manning, C.D.: GloVe: global vectors for word repre-
sentation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural
Language Processing, pp. 1532–1543 (2014)
20. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre-
sentations in vector space. arXiv preprint, arXiv:1301.3781 (2013)
21. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-
sentations of words and phrases and their compositionality. In: Advances in Neural
Information Processing Systems, vol. 26, pp. 3111–3119 (2013)
22. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with
subword information. arXiv preprint arXiv:1607.04606 (2016)
23. Arora, S., Liang, Y., Ma, T.: A simple but tough-to-beat baseline for sentence
embeddings. In: Proceedings of 5th International Conference on Learning Repre-
sentations (2017)
Estimation of Emotion Type
and Intensity in Japanese Tweets Using
Multi-task Deep Learning
1 Introduction
The rest of this paper is organized as follows. Related works are described in
Sect. 2. In Sect. 3, we explain how to construct our database of Japanese tweets
in detail. The multi-task deep learning framework for emotion estimation is
proposed in Sect. 4. After showing the experimental results in Sect. 5, we conclude
the paper and describe future direction in Sect. 6.
2 Related Work
Many methods to estimate emotional intensity have been proposed in recent
years. Himeno [6] proposed a method to estimate the emotional intensity of
tweets by combining features obtained from a model called the DeepMoji model
[7] and those obtained using multiple affect lexicons based on EmoInt [8]. The
DeepMoji model is trained through emoji predictions on a dataset of 1.2 billion
tweets with 64 common emoji labels and is a model with high performance in
tasks such as emotion analysis and sarcasm classification by transfer learning.
The workshop WASSA-2017 [1] held the first international competition on
estimating the intensity of emotion in tweets. Participants were required to esti-
mate the emotional strength of the tweet, and their accuracies were evaluated
by Pearson correlation coefficients. In the workshop, 22 teams took part in the
shared task, and commonly used features included word embeddings and sen-
tence representations learned by neural networks. The top eight systems also
made use of a substantially larger number of affect lexicons, or dictionaries,
to generate features. The best performing system was Prayas [9] with 0.747 of
Pearson correlation. The system used an ensemble of three different models:
feed-forward neural network, multi-task deep learning and sequence modeling
using Convolutional Neural Networks (CNN) [10] and Long Short-Term Memory
(LSTM) [11]. The YZU-NLP [13] proposed a combined model of Bidirectional
LSTM (BiLSTM) [12] and CNN without dictionaries to encode the global infor-
mation captured by LSTM among the most principal features extracted by CNN.
The system achieved 0.677 of correlation. As similar, the YNU-HPCC [14] with
0.671 of correlation also utilized a combined model of CNN and LSTM without
dictionaries to extract both local information within tweets and long-distance
dependency across tweets in the regression process.
Another shared task on emotional intensity estimation was held in the work-
shop SemEval2018 [2]. In the competition, SeerNet [15] won the first place by
316 K. Sato and T. Ozaki
All Tweets
Model 1: The YZU-NLP [13] treats a tweet as a word sequence and processes it
in the order of a Bi-LSTM layer, a Convolution layer, a Max-Pooling layer,
a Dense Layer. This model is trained by the mean squared error.
Model 2: The YNU-HPCC [14] decomposes tweets into word sequences and
processes them in the order of a Convolution layer, a Max-Pooling layer,
318 K. Sato and T. Ozaki
a LSTM layer, a Dense layer. Note that, compared with YZU-NLP, the
Convolution and LSTM layers are arranged in reverse order. YNU-HPCC
employs the mean absolute error in the training.
Model 3: This model [18] combines word level representation and character level
representation. While word level inputs are processed with a Bidirectional
Gated Recurrent Unit (Bi-GRU) [19], character level inputs are processed in
the order of a Convolution layer with multiple filter sizes and a GRU layer.
These two outputs are passed into a balancer named Highway Networks [20]
to output the prediction value. While this method is originally developed
for the classification tasks using the cross entropy with L2 regularization,
we modify it for the regression tasks with mean squared error.
5 Experiments
6 Conclusion
In this paper, to achieve accurate estimations of emotional types and intensi-
ties of Japanese tweets, we constructed a database of Japanese tweets anno-
tated for joy, anger, sadness and pleasant intensities using the best-worst scaling
method. In addition, we proposed a multi-task deep learning framework based on
three existing models. Through the experiments, we confirmed that the proposed
framework can improve the estimation accuracy on emotional intensities.
As future work, it is necessary to evaluate the proposed framework using large
scale datasets with more emotional types. It is also necessary to develop certain
mechanisms for balancing the classification and regression tasks. We also plan
to develop a model which combines the unsupervised representation learning for
tweets and the ensemble learning with multiple estimators.
References
1. Mohammad, S.M., Bravo-Marquez, F.: WASSA-2017 shared task on emotion inten-
sity. In: Proceedings of the 8th Workshop on Computational Approaches to Sub-
jectivity, Sentiment and Social Media Analysis (2017)
2. Mohammad, S.M., Bravo-Marquez, F., Salameh, M., Kiritchenko, S.: SemEval-
2018 task 1: affect in Tweets. In: Proceedings of the 12th International Workshop
on Semantic Evaluation (2018)
3. Go, A., Bhayani, R., Huang, L.: Twitter sentiment classification using distant
supervision. In: Final Projects from CS224N for Spring 2008/2009 at The Stanford
Natural Language Processing Group (2009)
4. Mohammad, S.M.: Word affect intensities. arXiv preprint arXiv:1704.08798 (2017)
322 K. Sato and T. Ozaki
20. Srivastava, R.K., Greff, K., Schmidhuber, J.: Training very deep networks. In:
Proceedings of the 28th International Conference on Neural Information Processing
Systems, vol. 2, pp. 2377–2385 (2015)
21. Kingma, D., Ba, J.: Adam: a method for stochastic optimization. arXiv preprint,
arXiv:1412.6980 (2014)
A Method for Extracting Influential People
for the Improvement of Contents
Abstract. It is very important to find influential users who have useful infor-
mation as well as extract helpful knowledge that affect the efficiency of user’s
intellectual work; however, it is difficult to evaluate how such helpful comments
and conversations from influential users can affect works of other users. In this
study, we propose a method for extracting user comments that influenced other
users to improve their contents such as a presentation slide by analyzing cor-
relation between user’s comments and improved parts of contents. Our method
allows us to make actual evaluation to influential users for some specific users.
In the experiment, we evaluate the feasibility of the proposed method using the
actual communication history that is collected from a communication service,
Slack.
1 Introduction
Information sharing services and collaborative developing services are widely used for
sharing the knowledge and information on intellectual and productive activities such as
software development and academic research. Using these services, users are doing
various activities such as discussion, work instruction, content share, and schedule
management for the purpose of knowledge and information sharing, and expertise
knowledge that users have is accumulated in various forms such as conversation
stream, document, presentation slide, and source code.
In this study, we focus on ‘influential users’ who often give advice and provide
useful information to some users in such activities of information sharing. By lever-
aging knowledge and information of influential users, it is deemed that the cost of
information sharing with other users can be reduced and the motivation to effectively
use information for some specific purpose can be improved. In our previous study, we
have proposed a method for extracting users who have influential roles, such as leader,
moderator, supporter, and facilitator, that are extracted from the technical communi-
cation history for extracting influential users for other users in a communication group
[1]. In this method, the frequency of messages and the number of users who are
communicated with are considered for extracting user’s roles, but how the knowledge
and information of the influential users who have special roles can affect other users are
not considered.
It is deemed that the influential users have the characteristics of ‘having useful
expertise knowledge’, ‘having reliable information’, ‘having excellent capability of
communication and information sharing’, ‘having instructional skill and problem-
solving ability’, and so on. Therefore, when a user is seeking the desired information,
from the above reasons, there is a possibility that the user can efficiently obtain useful
and reliable information by extracting information that can be derived from users who
already build reliable relationships with others, and can continue the meaningful
communication with them. Thus, it is very important to find influential users who have
such useful information as well as extract helpful knowledge from accumulated
information [2–4], which affect the efficiency of user’s intellectual work; however, it is
difficult to evaluate how such helpful comments and conversations from influential
users can affect works of other users, even if they were actually effective for improving
the efficiency of user’s work and motivating users to take action according to their
information.
In this study, for evaluating how the knowledge and information from influential
users are used by other users on an online community of technical information sharing,
we propose a method for extracting user comments that influenced other users to
improve their contents such as a presentation slide by analyzing correlation between
user’s comments and improved parts of contents. Our method allows us to make actual
evaluation to influential users for some specific users.
In the experiment, we evaluate the feasibility of the proposed method using actual
data set up by collecting presentation slides which are updated every week for research
discussion in a laboratory and user’s comments through the discussion for improving
their slides and researches.
2 Proposed Method
Figure 1 shows an overview of the proposed method. First, our method calculates
correlation between user’s comments and difference of contents updated by other users.
If correlation value is high, we can assume that the content might be improved based on
the user’s comments. Based on this assumption, our method extracts the user comments
that influenced some specific users to improve their contents and regards a user who
have such useful comments as an influential user.
Tables 1 and 2 show an example of user comments and an example difference of
sentences between old slide and new slide, respectively. As shown in Table 1, com-
ments from each user consists of several sentences for the purpose of improving one
presentation slide. In addition, Table 2 shows that there are five updated parts of
sentences in a presentation slide based on user’s comments.
326 H. Tsukiji and K. Takano
User
Extract the difference of content and content
updated in a certain period
Content C1 Content C1
Comment msg3
Comment msg1
User C
User A Comment msg6
User B Comment msg2
User D
User C Comment msg3 Comment msg4
User D
…
…
User B
Get comments from other users in a certain period By calculating correlation between comment
and difference of updated content, influential
comments are extracted.
…
Ci Ci Ci … Ci
Other users
u2 un
MSG(Ci)
…
ΔC1 ΔC2 … ΔCl-1
MSG(Cn)
…
Sentence dm
MSGn(Ci, Δt1)
Rank users who is the most influen al for
Extract influen al comments for other users by calcula ng similarity score user u1 based on the sum total of similarity
between difference ΔC and comment MSG(Ci) score in a certain period t
3 Experiments
In the experiment, we used the actual data set by collecting presentation slides which
are updated every week for research discussion in a laboratory and user’s comments
through the discussion for improving their slides and researches. Table 3 shows the
number of experimenters. Three users User1, User2, and User3 only gave their com-
ments and eight users User4 to User11 updated their slides as well as gave comments to
other users. In addition, Tables 4 and 5 show the number of difference of sentences in
slides and the number of comments received from other users, respectively.
From eight users, we focus on the result of User8 and User10 whose frequency of
updating slide is high. Tables 6, 7, and 8 show user comments for User8, difference of
sentences in slide of User8, and correlation values between user comments and dif-
ference of sentences by User8, respectively. Similarly, Tables 9, 10, and 11 show the
result of User10. From Table 8, we can find that correlation values for User1, User2,
and User10 are high, and this means User8 improves the slides based on these three
users. From these results, it is deemed that comments from User1, User2, and User 10
are influential for User8. Similarly, from Tables 9, 10, and 11, we can confirm that
comments from User1, User3, and User6 are influential for User10.
Table 8. Correlation values between user comments and difference of sentences by User8
Id User1 User2 User4 User10
4 0.189 0.000 0.000 0.000
67 0.157 0.040 0.133 0.135
95 0.089 0.079 0.067 0.135
101 0.058 0.000 0.086 0.174
213 0.121 0.160 0.135 0.091
Table 11. Correlation values between user comments and difference of sentences by User10
Id User1 User2 User3 User6 User7 User8 User9
22 0.177 0.000 0.139 0.267 0.000 0.000 0.000
23 0.306 0.091 0.240 0.154 0.000 0.183 0.115
29 0.395 0.071 0.248 0.239 0.169 0.141 0.089
In addition, Tables 12, 13, 14, and 15 shows the ranking results of total sum of
correlation values for four users: User4, User8, User10, and User11, respectively.
In Table 8, User2, User1, and User8 are ranked higher. This is because User4 has
improved slides according to their comments, and we can see that User2, User1, and
User8 are influential users for User4. Similarly, it is deemed that User2 and User1 are
influential for User4, User2, User3, and User1 are influential for User10, User4 and
User2 are influential for User 11.
From these result, we can confirm that proposed method can extract influential
users based on correlation between user’s comments and improved parts of contents.
4 Conclusion
This paper has presented a method for extracting user comments that influenced other
users to improve their contents by analyzing correlation between user’s comments and
improved parts of contents. In the experiment, we evaluated the feasibility of the
proposed method using the actual communication history that is collected from a
communication service, Slack, and confirmed that our method can extract influential
users by calculating correlation values between user’s comments and improved parts of
contents.
In our future work, we will evaluate the effectiveness of our method using larger
size of communication history in longer period. In addition, we will develop a rec-
ommendation service to provide useful knowledge and information of influential users
by applying our method to recommendation algorithm.
References
1. Hayato, T., Kosuke, T.: An estimation method of user roles based on online technical
communication activities. In: Proceedings of the Workshops of the 32nd International
Conference on Advanced Information Networking and Applications (WAINA-2018),
pp. 181–186 (2018)
2. Khan, H.U., Daud, A., Malik, T.A.: MIIB a metric to identify top influential bloggers in a
community. PLoS ONE 10(9), e0138359 (2015)
3. Akritidis, L., Katsaros, D., Bozanis, P.: Identifying the productive and influential bloggers in
a community. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 41(5), 759–764 (2011)
4. Ishfaq, U., Khan, H.U., Khalid, I.: Identifying the influential bloggers a modular approach
based on sentiment analysis. J. Web Eng. 16(5&6), 505–523 (2017)
5. Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P, Riedl, J.: GroupLens an open
architecture for collaborative filtering of netnews. In: CSCW 1994, Proceedings of the 1994
ACM Conference on Computer Supported Cooperative Work, pp. 175–186 (1994)
6. Truong, K.Q., Ishikawa, F., Honiden, S.: Improving accuracy of recommender system by
item clustering. IEICE Trans. Inf. Syst. 90(9), 1363–1373 (2007)
7. Tan, L.K.W., Na, J.C., Theng, Y.L.: Influence detection between blog posts through blog
features, content analysis, and community identity. Online Inf. Rev. 35(3), 425–442 (2011)
8. Zhang, S., Zhang, S., Yen, N.Y., Zhu, G.: The recommendation system of micro-blog topic
based on user clustering. Mobile Netw. Appl. 22(2), 228–239 (2017)
9. Li, W., Ye, Z., Xin, M., Jin, Q.: Social recommendation based on trust and influence in SNS
environments. Multimed. Tools Appl. 76(9), 11585–11602 (2017)
A Method for Extracting Influential People for the Improvement of Contents 333
10. Seo, Y.-D., Kim, Y.-G., Lee, E., Baik, D.-K.: Personalized recommender system based on
friendship strength in social network services. Expert Syst. Appl. 69(1), 135–148 (2017)
11. Basher, A.R.M.A., Fung, B.C.M.: Analyzing topics and authors in chat logs for crime
investigation. Knowl. Inf. Syst. 39(2), 351–381 (2014)
12. Yasin, M., Qureshi, J.A., Kausar, F., Kim, J.: A granular approach for user-centric network
analysis to identify digital evidence. Peer Peer Netw. Appl. 8(5), 911–924 (2015)
13. Wang, T., Huang, Z., Gan, C.: On mining latent topics from healthcare chat logs. J. Biomed.
Inform. 61, 247–259 (2016)
Sentiment Analysis of Arabic
and English Tweets
Abstract. Due to the continuous and rapid growth of daily posted data on the
social media sites in many different languages, the automated classification of
this huge amount of data has become one of the most important tasks for
handling, managing, and organizing this huge amount of textual data. There
exist many examples of social media sites, but Twitter is considered to be one of
the most popular and commonly used, as users are able to communicate with
each other, share their opinions, and express their emotions (sentiments) in the
form of convenient short blogs using less than 140 words. Accordingly, many
companies and organizations may analyze these sentiments in order to evaluate
the users’ thoughts, and determine their polarity from the content of the text. For
this process, natural language processing techniques, statistics, or machine
learning algorithms are being used to identify and extract the sentiment of the
text. In practice, many data mining techniques and algorithms are being applied
to observe patterns and correlation among that huge amount of data. This paper
proposes an efficient approach in handling Tweets, in both Arabic and English
languages, with different processing techniques applied. This approach is based
on using the Vector Space Model (VSM) to represent text documents and
Tweets, and the Term Frequency Inverse Document Frequency (TFIDF) in a
term weighting process to generate the feature vector for classification process.
The proposed approach has been evaluated using several experiments with
different classifiers on five datasets: Decision trees, Naive-Bayes, kNN, Logistic
Regression, Perceptron, and Multilayer Perceptron. The experimental results
reveal the effectiveness of our proposed approach when comparing classification
results with the published work in [1–3].
1 Introduction
In the past twenty years, social network sites have made the entire world connected
together. Using these networks, users are able to share many things, such as views,
feelings, experiences, advice, etc. These electronic Word of Mouth (eWOM) statements
expressed on the web could be beneficial and helpful to others [4, 5]. There are many
social network sites but Twitter is one of the most popular and commonly used social
network sites, where users can easily communicate with each other or share their
opinions in the form of short blogs less than 140 words. The huge amount of posted
text on Twitter is increasing in an astronomical pace. This rapid growth in data makes it
difficult to handle especially in a real-time environment with existing databases and
techniques.
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 334–348, 2019.
https://doi.org/10.1007/978-3-030-15035-8_32
Sentiment Analysis of Arabic and English Tweets 335
Sentiment analysis has become one of the most widely researched topics in dif-
ferent aspects. The word sentiment refers to the feeling, thought, or opinion, of a person
across some particular domains. Hence, sentiment analysis is also referred to as opinion
mining. The huge amount of Tweets posted daily makes Twitter data one of the most
essential data volumes that deal with business, industry or social issues. In fact, as the
posted twitter data satisfies the properties that it is huge in size, comes from hetero-
geneous sources, is in both structured and unstructured forms, has high speed in data
generation, and is very inconsistent, so it could be considered as big data [6]. The main
goal behind the analysis of sentiments, is to evaluate these thoughts, determine its
polarity using NLP (Natural Language Processing), tagging, statistics, or machine
learning algorithms to identify and extract the sentiment content of a text.
Technically, data mining’s main goal is to observe correlations or patterns among a
huge amount of data in large relational databases. Nowadays it is not necessary to turn
to large structured databases to have large volume of data analyzed [7]. The traditional
way in acquiring knowledge from data relies on the manual analysis and interpretation
carried out by human experts who are familiar with the data domain. They serve as an
interface between the data and the users. Unfortunately, this manual approach is slow,
expensive, and highly subjective, as well as being dependent on the experience of the
experts, and it has become impracticable due to the dramatically increased volume of
data on the web [8].
In this work, the automated handling of Tweets, in both Arabic and English lan-
guages, was investigated and different processing techniques were applied on five
different datasets of various sizes in both languages. Section 2 shows the related work.
Section 3 presents the proposed multi-stage model for sentiment analysis, and shows
the details of each stage. Experimental setup and discussion are given in Sect. 4.
Section 5 summarizes the overall conclusion and expected directions for future work.
2 Related Work
Sentiment Analysis is the area of web data mining which deals with evaluating the
feelings or thoughts of people across some particular domains, and to determine the
polarity and to what degree [9]. In the past few years, the field of detecting and
evaluating the feelings of users or customers in many domains has been extensively
investigated. This is due to the growth of data which exists online over the internet,
especially those that deal with sentiment analysis such as describing people’s point of
view, experiences, preferences, and thoughts.
Neethu and Rajasree [10] and Jindal et al. [11] gave detailed and deep reviews on
machine learning approaches, documents representation techniques, datasets, and
techniques used for text-documents classification. It is clearly indicated from the
reviews that the Vector-Space-Model (VSM) is the most commonly used method for
document and text classification, with each document being represented as a compact
set of words. Ravi et al. [9] introduced a survey that covers the different types of
sentiment analysis, datasets used, natural language processing techniques used, and
their applications, as well as explanation and categorization of some sentiment analysis
algorithms and their originating references. Also, Yang et al. [12] introduced the
336 M. K. Elhadad et al.
Training Testing
Pre-Processing Pre-Processing
Classification
Classification Stage
Algorithm
Learning Stage
Selection Apply
Classification
Process
Building the
Classification
Model
Output:
Classified Documents/Tweets
In the Training phase, the two stages are Feature Extraction and Learning while in
the Testing phase, the stages are Feature Extraction and Classification. Feature
Selection is described in Subsect. 3.1 and Learning and Classification are presented in
Subsect. 3.2. Though the model is designed generically for any language, in our current
work, we are dealing with Arabic and English languages only.
While computing TF, all terms are considered equally important [31]. IDF mea-
sures how important a term is,
Total Number of Tweets
IDFðtÞ ¼ log
Number of Tweets that contains term (t)
Finally, the TF-IDF measure the multiplication of TF and IDF, as follows [31],
This means that larger weights are assigned to terms that appear relatively infre-
quent throughout the corpus, but very frequently in individual document/Tweets.
In this section, we discuss our experimental setup and the results for evaluating the
performance of our proposed approach on five different datasets. There are two datasets
in Arabic: the Arabic Sentiment Tweets Dataset (ASTD) [2] and the Restaurant
Reviews Dataset (RES) [33]. For English, there are three datasets: the Stanford Twitter
sentiment dataset [34], the Twitter US Airline Sentiment dataset [35] and the Uber Ride
Reviews dataset [36].
• The Restaurant Reviews Dataset (RES): The restaurants domain data is taken from
two sources. The first is “Qaym4” with 8.6K Arabic reviews, and the second is
collected from “TripAdvisor” and has 2.6K reviews. Both datasets cover 4.5K
restaurants and have reviews written by over 3 K users.
• Arabic Sentiment Tweets Dataset (ASTD): It contains over 10K Arabic Sentiment
Tweets classified into four classes: subjective positive, subjective negative, sub-
jective mixed, and objective.
• The Stanford Twitter sentiment corpus (Stanford): It contains 1.6 million Tweets
automatically labelled as positive or negative based on emotions.
• The Twitter US Airline Sentiment Dataset (US Airline): This data originally came
from Crowdflower Data for Everyone library. It was scraped from February of 2015
and it has been classified by the contributors as positive, negative, and neutral
Tweets.
• The Uber Ride Reviews Dataset (Uber): It contains 1345 Uber ride reviews that are
collected in the period between 2014–2017.
2 Precision Recall
F1-measure (Precision, Recall) =
Precision + Recall
In the above formula, precision, and recall are two standard measures widely used
in text categorization literature to evaluate an algorithm’s effectiveness on a given
category where:
True Positive
Precision =
True Positive + False Positive
342 M. K. Elhadad et al.
True Positive
Recall =
True Positive + False Negative
In addition, it should be noted that the accuracy of the Bernoulli Naive Bayes
classifier when applied to the US Airline Tweets dataset (76%) and the Uber Ride
Reviews dataset (84%), performs much better than the other classifiers. Furthermore,
the accuracy of Logistic Regression classifier (79%) when applied to the Stanford
Tweets dataset (79%), is higher than the other classifiers.
Moreover, we noted that from the results obtained when applying the kNN clas-
sifier the following:
• The kNN classifier assigns equal weights to each attribute in the feature vector. This
may cause confusion when there are many irrelevant attributes in the data.
• Second, the traditional kNN classifier is based on measuring the distance between
the documents/Tweets using different distance measures, such as Euclidean and
Minkowski, but ignoring the semantic relations between them. This may lead to
misclassification of some of the tested data.
Furthermore, when comparing our obtained results with other results from [1–3],
we find the superiority of our proposed model in all seven classifiers as shown in
Table 1.
344 M. K. Elhadad et al.
This paper addresses the task of performing automatic sentiment analysis over textual
datasets in both Arabic and English languages. To achieve this goal, a classification
model is introduced where textual documents/Tweets pass through a two-stage process.
In the pre-processing step, NLP techniques are applied to get the BoWs that best
represents the datasets, and a textual feature vector is generated. Additionally, TF-IDF
is used to generate the numerical feature vector for classification tasks. After building
Sentiment Analysis of Arabic and English Tweets 345
the feature vector, different classifiers are used to test various models and their per-
formance. Performance is reported using different metrics: Accuracy, Precision,
F-measure, and Recall. Experimental results through the entire paper indicate that:
• For the RES dataset, the best result is obtained by the Multilayer Perceptron
algorithm with an accuracy of 81%. Also, in terms of true positive rate (TPR), the
best results are obtained by the same classifier with a recall of 82%.
• For the ASTD dataset, the best result was obtained using the Logistic Regression
algorithm with an accuracy of 68%. In terms of TPR, the best results are obtained
by the same classifier with a recall of 68%.
• For the US Airline Tweets dataset, the best result was obtained by the Bernoulli
Naive Bayes algorithm with an accuracy of 76%. But in terms of TPR, the best
result is obtained by the Logistic Regression classifier with a recall of 74%.
• For the Uber Ride Reviews dataset, the best result was obtained by both the Ber-
noulli Naive Bayes algorithm and the kNN classification algorithm with an accu-
racy of 84%. In terms of TPR, the best result was obtained by the same classifiers
with a recall of 84%.
• For the Stanford Tweets dataset, the best result was obtained by Logistic Regression
with an accuracy of 79%. In terms of TPR, the best result was obtained by the same
classifier with a recall of 79%.
Note that kNN and the Multilayer Perceptron classification algorithms do not have
any scores. This is due to hardware limitations.
For future work, there are many avenues to pursue, including:
• Utilize the WordNet ontology for building the extracted feature vector to improve
the classification accuracy.
• Extend and evaluate the proposed system against other datasets in other languages
other than Arabic or English.
• Apply dimensionality reduction techniques to minimize the size of the used feature
vector without affecting the overall system performance.
• Extend the proposed work to handle Tweets that uses Franco-Arabic texts.
• Improve the existing techniques and methodologies to be used in classifying web
text documents.
• Implement a web text document classification system using parallel platforms, such
as Apache Spark, to handle the large amount of data in less time without affecting
the overall system performance.
• Combine the decisions of two or more classification algorithms to improve the
accuracy of the results.
• Perform clustering tasks over the set of feature vectors to help application to give
accurate results and search quickly.
346 M. K. Elhadad et al.
References
1. Rane, A., Anand, K.: Sentiment classification system of Twitter data for US airline service
analysis. In: 2018 IEEE 42nd Annual Computer Software and Applications Conference
(COMPSAC), pp. 769–773. IEEE, Tokyo (2018)
2. Nabil, M., Alay, M., Atiya, A.: ASTD: Arabic sentiment tweets dataset. In: Proceedings of
the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 2515–
2519. Association for Computational Linguistics, Lisbon (2015)
3. Baly, R., Badaro, G., El-Khoury, G., Moukalled, R., Aoun, R., Hajj, H., El-Hajj, W.,
Habash, N., Shaban, K.: A characterization study of Arabic Twitter data with a
benchmarking for state-of-the-art opinion mining models. In: Proceedings of the Third
Arabic Natural Language Processing Workshop (WANLP), pp. 110–118. Association for
Computational Linguistics, Valencia (2017)
4. Heikal, M., Torki, M., El-Makky, N.: Sentiment analysis of Arabic Tweets using deep
learning. Procedia Comput. Sci. 142, 114–122 (2018)
5. Goel, A., Gautam, J., Kumar, S.: Real time sentiment analysis of Tweets using Naive Bayes.
In: 2016 2nd International Conference on Next Generation Computing Technologies
(NGCT). IEEE (2016)
6. Holzinger, A., Stocker, C., Ofner, B., Prohaska, G., Brabenetz, A., Hofmann-Wellenhof, R.:
Combining HCI. In: Natural Language Processing, and Knowledge Discovery-Potential of
IBM Content Analytics as an Assistive Technology in the Biomedical Field, pp. 13–24.
Springer, Heidelberg (2013)
7. Ristoski, P., Paulheim, H.: Semantic web in data mining and knowledge discovery: a
comprehensive survey. Web Semant. Sci. Serv. Agents World Wide Web 36, 1–22 (2016)
8. Mukherjee, S., Shaw, R., Haldar, N., Changdar, S.: A survey of data mining applications and
techniques. Int. J. Comput. Sci. Inf. Technol. (IJCSIT) 6(5), 4663–4666 (2015)
9. Ravi, K., Ravi, V.: A survey on opinion mining and sentiment analysis: tasks, approaches
and applications. Knowl. Based Syst. 89, 14–46 (2015)
10. Neethu, M.S., Rajasree, R.: Sentiment analysis in Twitter using machine learning
techniques. In: 2013 Fourth International Conference on Computing, Communications and
Networking Technologies (ICCCNT), pp. 1–5. IEEE, Tiruchengode (2013)
11. Jindal, R., Malhotra, R., Jain, A.: Techniques for text classification: literature review and
current trends. Webology 12(2), 1–28 (2015)
12. Yang, P., Chen, Y.: A survey on sentiment analysis by using machine learning methods. In:
2017 IEEE 2nd Information Technology, Networking, Electronic and Automation Control
Conference (ITNEC), pp. 117–121. IEEE, Chengdu (2017)
13. Pang, B., Lee, L., Vaithyanathan, S.: Thumbs up? Sentiment classification using machine
learning techniques. In: Proceedings of the ACL-02 Conference on Empirical Methods in
Natural Language Processing (EMNLP), pp. 79–86. Association for Computational
Linguistics, ACM, Philadelphia (2002)
14. Das, S., Behera, R.K., Rath, S.K.: Real-time sentiment analysis of Twitter streaming data for
stock prediction. Procedia Comput. Sci. 132, 956–964 (2018)
15. Kim, S.-B., Han, K.-S., Rim, H.-C., Myaeng, S.H.: Some effective techniques for Naive
Bayes text classification. IEEE Trans. Knowl. Data Eng. 18(11), 1457–1466 (2006)
16. Niu, Z., Yin, Z., Kong, X.: Sentiment classification for microblog by machine learning. In:
2012 Fourth International Conference on Computational and Information Sciences, pp. 286–
289. IEEE, Chongqing (2012)
Sentiment Analysis of Arabic and English Tweets 347
17. Barbosa, L., Feng, J.: Robust sentiment detection on Twitter from biased and noisy data. In:
Proceedings of the 23rd International Conference on Computational Linguistics: Posters,
COLING 2010, pp. 36–44. ACM, Beijing (2010)
18. Celikyilmaz, A., Hakkani-Tür, D., Feng, J.: Probabilistic model-based sentiment analysis of
Twitter messages. In: 2010 IEEE Spoken Language Technology Workshop, pp. 79–84.
IEEE, Berkeley (2011)
19. Guellil, I., Adeel, A., Azouaou, F., Hachani, A.E., Hussain, A.: Arabizi sentiment analysis
based on transliteration and automatic corpus annotation. In: Proceedings of the 9th
Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media
Analysis, pp. 335–341. Association for Computational Linguistics, Brussels (2018)
20. Bhavitha, B.K., Rodrigues, A.P., Chiplunkar, N.J.: Comparative study of machine learning
techniques in sentimental analysis. In: 2017 International Conference on Inventive
Communication and Computational Technologies (ICICCT), pp. 216–221. IEEE, Coimbat-
ore (2017)
21. Dashtipour, K., Poria, S., Hussain, A., Cambria, E., Hawalah, A.Y., Gelbukh, A., Zhou, Q.:
Multilingual sentiment analysis: state of the art and independent comparison of techniques.
Cogn. Comput. 8(4), 757–771 (2016)
22. Balahur, A., Turchi, M.: Multilingual sentiment analysis using machine translation? In:
Proceedings of the 3rd Workshop on Computational Approaches to Subjectivity and
Sentiment Analysis, pp. 52–60. Association for Computational Linguistics, Jeju (2012)
23. Sedding, J., Kazakov, D.: WordNet-based text document clustering. In: Proceedings of the
3rd Workshop on Robust Methods in Analysis of Natural Language Data, pp. 104–113.
Association for Computational Linguistics, ACM, Geneva (2004)
24. Peng, X., Choi, B.: Document classifications based on word semantic hierarchies. Artif.
Intell. Appl. 5, 362–367 (2005)
25. Marcus, M.P., Marcinkiewics, M.A., Santorini, B.: Building a large annotated corpus of
English: the Penn Treebank. Comput. Linguist. 19(2), 313–330 (1993). Special issue on
using large corpora: II (Association for Computational Linguistics, MIT Press Cambridge)
26. Maamouri, M., Bies, A., Krouna, S., Gaddeche, F., Bouziri, B.: Arabic tree banking
morphological analysis and pos annotation, Ver. 3.8. Linguistic Data Consortium, Univ.
Pennsylvania, Pennsylvania (2009)
27. Albared, M., Omar, N., Ab Aziz, M.J., Nazri, M.Z.A.: Automatic part of speech tagging for
Arabic: an experiment using Bigram hidden Markov model. In: International Conference on
Rough Sets and Knowledge Technology, pp. 361–370. Springer, Heidelberg (2010)
28. NL company, Ranks: Stopword Lists. ranks.nl. (n.d). https://www.ranks.nl/stopwords.
Accessed 19 Jan 2019
29. Wen, C.Y.J.: Text categorization based on a similarity approach. In: Proceedings of
International Conference on Intelligent System and Knowledge Engineering, pp. 1–5.
Atlantis Press, China (2007)
30. Wu, H.C., Luk, R.W.P., Wong, K.F., Kwok, L.: Interpreting TF-IDF term weights as making
relevance decisions. ACM Trans. Inf. Syst. 26(3), 13 (2008)
31. Schütze, H., Manning, C.D., Raghavan, P.: Introduction to Information Retrieval, vol. 39.
Cambridge University Press, Cambridge (2008)
32. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J.: Scikit-learn: machine learning in
Python. J. Mach. Learn. Res. 12, 2825–2830 (2011)
33. ElSahar, H., El-Beltagy, S.R.: Building large arabic multi-domain resources for sentiment
analysis. In: Gelbukh, A. (ed.) Computational Linguistics and Intelligent Text Processing,
CICLing 2015. LNCS, pp. 23–34. Springer (2015)
348 M. K. Elhadad et al.
34. Go, A., Bhayani, R., Huang, L.: Twitter sentiment classification using distant supervision,
p. 12. CS224N Project Report, Stanford (2009)
35. Twitter-Airline-Sentiment Dataset: taken from the standard Kaggle, vol. 2. Kaggle (2017)
36. Purvank. Uber-rider-reviews-dataset taken from the standard Kaggle, kaggle (2018)
37. Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for classification
tasks. Inf. Process. Manage. 45(4), 427–437 (2009)
The 12th International Workshop on Bio
and Intelligent Computing
(BICom-2019)
A Deep Q-Network with Experience
Optimization (DQN-EO) for Atari’s
Space Invaders
1 Introduction
In this paper we use the Atari environment, and apply our proposed algorithm
in the popular game of “Space Invaders”. The game consists of a ground unit,
which is controlled by the user, which tries to eliminate the spaceships above it,
by shooting laser beams only upwards. The spaceships move sideways, until they
reach the side of the screen, where they step closer to the user-controlled ground
unit. While moving, they also shoot laser beams downward, so the user should
avoid two situations; being hit by the laser beams and letting the spaceships reach
the ground. The user scores points for each eliminated spaceship, depending on
the position of the spaceship (the highest the ship, the highest the score). If the
ground unit is eliminated three times, the game is over, and the accumulated
score is the final result of the game.
In this work, we propose a Deep Q-Network with Experience Optimization
(DQN-EO) algorithm. We use the Space Invaders’ (OpenAI’s Gym) environment
and apply our algorithm in order to evaluate its performance, regarding reward
drop rate, total episode score and episode lifetime. The remainder of the paper
is organized as follows.
In Sect. 2, we summarize the basics of Reinforcement learning, and how it
is applied to the Space Invaders environment, which is presented in Sect. 3. In
Sect. 4, we describe the four phases of our proposed DQN-EO algorithm; experi-
ence optimization, search algorithms and neural network structure. Training of
the algorithm is explained in Sect. 5. Then, we discuss the results in Sect. 6 and
conclude our paper in Sect. 7.
2 Reinforcement Learning
In Reinforcement Learning, the agent gets a state s from environment env and
selects an action a to apply to the environment env. The environment will feed-
back a reward (Qπ(s, a)) to the agent by evaluating the action and its current
state (see Fig. 1). In the future the agent will be oriented to choose the action
which previously obtained the highest reward. Different reinforcement learning
algorithms have been developed and applied to different problems. To name a
few, Q-Learning, DQN, SARSA, DDPG are widely-used ones. In Q-Learning,
agents use Q-tables to store the Q-value for every action-state pair, that it ever
experienced. If agent gets a reward r in time t, Q-Learning will update Q-table
so that st will get the full reward r, st−1 , st−2 . . . st−n will get smaller rewards as
defined by the discount factor γ, respectively, r × γ, r × γ 2 . . . r × γ n . Q-Learning
solves simple problems quickly, but it also falls into an infinite loop easily. In
complicated problems, the number of possible states is so big, that it is not very
practical to store all of them in a Q table, because searching for an existing state
in a huge Q-table will consume valuable resources. Deep Q-Network (DQN) uses
Neural Networks together with Q-Learning, to solve this problem. DQN uses
states as Neural Networks’ input, with several hidden layers. Neural Networks
will calculate the reward Q for all possible actions and choose to apply the action
with the highest reward to the environment. However, feedback about an action
can only be received after several time-steps of the running algorithm, and the
A Deep Q-Network with Experience Optimization (DQN-EO) 353
effect of key actions (fire, move, jump) to the reward (score) is delayed and not
accurate. We will try to solve this problem in this paper.
3 Environment
Terminology Description
Agent – The entity that takes action a (see below)
on the environment env (see below)
Action (a) – Chosen from a list of possible actions,
depending on application
Environment (env) – The “world” in which the agent moves,
takes action and gets rewarded or penalized
State (s) – A location, situation or condition of the
agent in a certain time
Reward (r) – The feedback of the environment for
successful agent’s actions. A penalty when
the action fails
Discount factor (γ) – A factor multiplied to future rewards,
which makes future rewards worth less than
immediate rewards
Policy (π) – A strategy, which maps states to actions,
in order for the agent to determine its next
action based on the current state
Value (V π(s)) – The expected future reward multiplied by
the discount factor (see above). It is defined
as the expected reward of the current state
under policy (π)
Q-value or action value – Similar to value v, but it takes action a as
(Qπ(s, a)) a parameter
Original Action 3 3 3 3 5 5 5 5 4 4 4 4 2 2 2 2
Reduced Action 3 0 0 0 5 2 2 2 4 1 1 1 2 2 2 2
Search Space b × a × c
We build 2 types of networks for training our data. One network is similar to that
of NIPS 2013 (Net13). The input to the neural network consists of a 84 × 84 × 4
image produced from the experience base D. In the first hidden layer 16 filters
are used. Each of them has 8 × 8 size with stride 4 with the input image and
applies a nonlinear rectifier. The second hidden layer uses 32 filters. Each of them
has a 4 × 4 size with stride 2, again followed by a nonlinear rectifier. The final
hidden layer is fully-connected and consists of 256 rectifier units. The output
layer is a fully-connected linear layer with a single output for each valid action
(6 outputs).
The other one is the same as the network in Nature 2015 (Net15). The
input to the neural network consists of a 84 × 84 × 4 image produced from the
experience base D. In the first hidden layer 16 filters are used. Each of them has
8 × 8 size with stride 4 with the input image and applies a nonlinear rectifier. The
second hidden layer uses 32 filters. Each of them has a 4 × 4 size with stride 2,
again followed by a nonlinear rectifier. This is followed by a third convolutional
layer that uses 64 filters of 3 × 3 with stride 1 followed by a rectifier. The final
hidden layer is fully-connected and consists of 512 rectifier units. The output
layer is a fully-connected linear layer with a single output for each valid action
(6 outputs).
Fig. 2. Relation between the actual frame number x in the episode and the approx-
imation of the number of frames y that passed since the fire action until each score
358 Y. Chen and E. Kulla
5 Training
We will train 2 algorithms (the original from NIPS 2013 and the proposed DQN-
OE) with each Network. As rectifier, which is also known as activation function,
we use ReLU (Rectified Linear Unit), as shown in Figs. 3 and 4.
For simplicity, and without deteriorating the performance, during training,
we changed all images to 84 × 84 grayscale, all positive rewards to 1, And all
negative rewards to −1. In this case, Algorithm 2 will receive negative reward
every time the number of lives decreases, 3 times more than Algorithm 1, which
will only receive negative reward when the game is over (number of lives =
0). So, We added the negative reward in Algorithm 1 for fair comparison. We
used AdamOptimizer provided by TensorFlow with minibatches of size 32 and
an -greedy policy. value was initialized to 0.9, and programmed to reduce to
0.1 linearly over 1 million training steps. It will be fixed at 0.1 after 1 million
steps. We trained for a total of 60 million steps and used a replay memory
of 100 thousand steps. The algorithm selects an action every 4 frames, and
emulator will repeat this action in the next 3 frames. Training program will be
active every 4 frames, similar to action selector. We use the actual frame and
3 previous frames as input for the networks. Thus, 1 step actually considers 4
frames in total. The trainer will output model files every 10 thousand steps for
a total of 599 models after training. When trainer successfully outputs a model
or the reward is dropped, the algorithm will record the reward drop rate and
the reward, which are used for evaluation purposes.
6 Result
The Drop Rate, shows how often a reward is not counted towards training the
systems, because the search algorithm could not find a fire action in the buffer.
In Fig. 5, we can see that at the beginning there are 2.5% drops, but after 2.5
million train steps, the drop rate reduces to under 1%.
6.2 Score
The score shows, the total score that the agent got from each episode. In Fig. 6,
the average and maximum score for 4 different cases. We can notice that, in the
beginning of the training, DQN-EO got much higher average score and higher
maximum score than the original one. However when decreased to 0.1 at 1
million steps, DQN-EO’s performance decreased. In Net13, DQN-EO caught up
original agent at step 3 million and then had the same performance with the
original one. In Net15, DQN-EO was better after 2.5 million steps. Moreover it
got the highest average score (457) and highest score (1275) from all the cases.
360 Y. Chen and E. Kulla
6.3 Lifetime
In Fig. 7, lifetime of DQN-EO was higher than the original one. Higher lifetime
shows that agent with Experience Optimization was better in avoiding enemies’
laser.
7 Conclusions
References
1. Bellemare, M.G., Naddaf, Y., Veness, J., Bowling, M.: The arcade learning environ-
ment: an evaluation platform for general agents. J. Artif. Intell. Res. 47, 253–279
(2013)
2. OpenAI Gym Toolkit. https://gym.openai.com/
3. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
Riedmiller, M.: Playing Atari with Deep Reinforcement Learning. arXiv, 19 Decem-
ber 2013. https://arxiv.org/pdf/1312.5602v1.pdf
4. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie,
C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., Has-
sabis, D.: Human-level control through deep reinforcement learning. Nature. 518,
529–33 (2015). https://doi.org/10.1038/nature14236
Design of a Deep Q-Network Based
Simulation System for Actuation Decision
in Ambient Intelligence
1 Introduction
Ambient Intelligence (AmI) is the vision that technology will become invisible,
embedded in our natural surroundings, present whenever we need it, enabled
by simple and effortless interactions, attuned to all our senses, adaptive to users
and context and autonomously acting [1,2]. High quality information and content
must be available to any user, anywhere, at any time, and on any device.
In order that Ambient Intelligence becomes a reality, it should completely
envelope humans, without constraining them. Distributed embedded systems for
Ambient Intelligence are going to change the way we design embedded systems,
in general, as well as the way we think about such systems. But, more impor-
tantly, they will have a great impact on the way we live. Applications ranging
from safe driving systems, smart buildings and home security, smart fabrics or e-
textiles, to manufacturing systems and rescue and recovery operations in hostile
environments, are poised to become part of society and human lives.
There are a lot of works done on testbed for AmI. In [3], the authors present
a simulation environment that offers a library of Networked Control Systems
(NCS) blocks. Thus, the constraints can be considered and integrated in the
design process. They describe a real process, an inverted pendulum, which is
automated based on Mica nodes. These nodes were designed especially for Ambi-
ent Intelligence purposes. This real NCS serves as a challenging benchmark for
proving the AmI suitability of the controllers.
In [4], the authors present the development of an adaptive embedded agent,
based on a hybrid PCA-NFS scheme, able to perform true real-time control
of Ambient Intelligence environments in the long term. The proposed architec-
ture is a single-chip HW/SW architecture. It consists of a soft processor core
(SW partition), a set of NFS cores (HW partition), the HW/SW interface, and
input/output (I/O) peripherals. An application example based on data obtained
in an ubiquitous computing environment has been successfully implemented
using an FPGA of Xilinx’s Virtex 5 family [5].
In [6], the authors describe a framework to Context Acquisition Services and
Reasoning Algorithms (CASanDRA) to be directly consumed by any type of
application needing to handle context information. CASanDRA decouples the
acquisition and inference tasks from the application development by offering a
set of interfaces for information retrieval. The framework design is based on a
data fusion-oriented architecture. CASanDRA has been designed to be easily
scalable; it simplifies the integration of both new sensor access interfaces and
fusion algorithms deployment, as it also aims at serving as a testbed for research.
There are many critical issues arising in AmI. Based on the specific applica-
tion, different objectives can be taken into account such as energy consumption,
throughput, delay, coverage, etc. [7,8]. Also many schemes have been proposed
in order to optimize a specific Quality of Service (QoS) parameter. We propose
a simulation system based on Deep Q-Network (DQN) as a controller for actor
node mobility and implement by Rust programming language [9,10]. We describe
the implementation and show the simulation results of the simulation system.
The rest of the paper is organized as follows. In Sect. 2, we describe the basics
of Ambient Intelligence. In Sect. 3, we present the overview of DQN. In Sect. 4,
we show the description and design of the simulation system. Simulation results
are shown in Sect. 5. Finally, conclusions and future work are given in Sect. 6.
2 Ambient Intelligence
In the future, small devices will monitor the health status in a continuous man-
ner, diagnose any possible health conditions, have conversation with people to
persuade them to change the lifestyle for maintaining better health, and com-
municates with the doctor, if needed [11]. The device might even be embedded
into the regular clothing fibers in the form of very tiny sensors and it might
communicate with other devices including the variety of sensors embedded into
the home to monitor the lifestyle. For example, people might be alarmed about
the lack of a healthy diet based on the items present in the fridge and based on
what they are eating outside regularly.
The Ambient Intelligence paradigm represents the future vision of intelligent
computing where environments support the people inhabiting them [12–14]. In
this new computing paradigm, the conventional input and output media no
364 T. Oda et al.
longer exist, rather the sensors and processors will be integrated into everyday
objects, working together in harmony in order to support the inhabitants [15]. In
these environments, a multitude of interconnected, invisible embedded systems,
seamlessly integrated into the background, surround the user [16]. The system
programs to meet their needs by recognizing people living in it and learning
from action. By relying on various artificial intelligence techniques, Ambient
Intelligence promises the successful interpretation of the wealth of contextual
information obtained from such embedded sensors, and will adapt the environ-
ment to the user needs in a transparent and anticipatory manner.
As necessary for Ambient Intelligence, media management and processing,
natural interaction, computer intelligence, emotion calculation etc. are included.
Technology that can accurately judge surrounding conditions by machine learn-
ing or artificial intelligence is necessary. Thereby, it is possible to flexibly pro-
vide the most suitable service at that time. Artificial intelligence is important to
achieve AmI and it is thought that it is necessary to receive more information
from artificial intelligence [17].
3 DQN
The change of the game score rt was regarded as the reward for the action. For
a standard reinforcement learning method, we can complete all of these game
sequences st as Markov decision process directly, where sequences of actions
and observations st = x1 , a1 , x2 , . . . , at−1 , xt . Also, it uses a technique known as
experience replay [26] in which it store the agent’s experiences at each time-step,
et = (st , at , rt , st+1 ) in a data-set D = e1 , . . . , eN , pooled over many episodes
into a replay memory. Defining the discounted reward for the future Tby atfactor
−t
γ, the sum of the future reward until the end would be Rt = t =t γ rt .
T means the termination time-step of the game. After performing experience
replay, the agent selects and executes an action according to an ε-greedy policy.
Since using histories of arbitrary length as inputs to a neural network can be
difficult, Q-function of DQN instead works on fixed length representation of
histories produced by a function φ. The target was to maximize the action-value
function Q∗ (s, a) = maxπ E[Rt |st = s, at = a, π], where π is the strategy for
choosing of best action. From the Bellman equation, it is equal to maximize the
expected value of r + γQ∗ (s , a ), if the optimal value Q∗ (s , a ) of the sequence
at the next time step is known.
Q∗ (s , a ) = Es ∼ξ [r + γmax
Q∗ (s , a )|s, a] (1)
a
The actuator nodes consider RGB LED. In order to decide the reward func-
tion, we considered the Uniformity (U) and Working Hours (WH) parameters.
The reward function r is defined as follows:
5 × (U + W H) (if staten ≥ statemax )
r= (4)
−1 × (U + W H) (else) .
Performance Evaluation of an AmI Testbed for Improving QoL 367
Parameters Values
Number of episode 100000
Number of iteration 100
Number of hidden layers 3
Number of hidden units 15
Initial weight value Normal initialization
Activation function ReLU
Action selection probability (ε) 0.999 − (t/Number of episode)
(t = 0, 1, 2, . . ., Number of episode)
Learning rate (α) 0.04
Discount rate (γ) 0.9
Experience memory size 300 × 100
Batch size 32
The initial weights values are considered as Normal Initialization [31]. The
input layer is using the position of events and actor nodes, mobile actor node
patterns and total reward values in Experience Memory. The hidden layer is
connected with 256 rectifier units in Rectified Linear Units (ReLU) [32]. The
output Q-values are mobile actuator node patterns.
Types of sensing
Type Values
Body motion left, right, up, down, stationary
Emotion delight, anger, sorrow, pleasure
Types of actuation
Type Values
Uniformity up, down, stationary
Light color white, red, blue, green, lights]out
5 Simulation Results
The simulation parameters are shown in Tables 1 and 2. One episode has 200
iterations.
368 T. Oda et al.
30000
20000
10000
Reward
-10000
-20000
Best
Median
Worst
-30000
0 50 100 150 200
Number of Iteration
Fig. 2. Results of reward.
In Table 3, we show the simulation total reward of episodes. There are best,
median and worst episodes. The best episode shows that actuator nodes are good
affect for visual environment. In Fig. 2, we show the simulation results of reward
vs. number of iteration. The higher reward value means that the actuator node
can effect visual environment.
6 Conclusion
References
1. Lindwer, M., Marculescu, D., Basten, T., Zimmermann, R., Marculescu, R., Jung,
S., Cantatore, E.: Ambient intelligence visions and achievements: linking abstract
ideas to real-world concepts. In: Design, Automation and Test in Europe Confer-
ence and Exhibition, pp. 10–15 (2003)
2. Obukata, R., Cuka, M., Elmazi, D., Oda, T., Ikeda, M., Barolli, L.: Design and
evaluation of an ambient intelligence testbed for improving quality of life. Int. J.
Space-Based Situated Comput. 7(1), 8–15 (2017)
3. Gabel, O., Litz, L., Reif, M.: NCS testbed for ambient intelligence. In: IEEE Inter-
national Conference on Systems, Man and Cybernetics, vol. 1, pp. 115–120 (2005)
4. del Campo, I., Martinez, M.V., Echanobe, J., Basterretxea, K.: A hard-
ware/software embedded agent for realtime control of ambient-intelligence envi-
ronments. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE),
pp. 1–8 (2012)
5. Virtex 5 Family Overview, Xilinx Inc., San Jose, CA (2009)
6. Bernardos, A.M., Tarrio, P., Casar, J.R.: CASanDRA: a framework to provide con-
text acquisition services and reasoning algorithms for ambient intelligence applica-
tions. In: International Conference on Parallel and Distributed Computing, Appli-
cations and Technologies, pp. 372–377 (2009)
7. Costanzo, C., Loscri, V., Natalizio, E., Razafindralambo, T.: Nodes self-deployment
for coverage maximization in mobile robot networks using an evolving neural net-
work. Comput. Commun. 35(9), 1047–1055 (2012)
8. Li, Y., Li, H., Wang, Y.: Neural-based control of a mobile robot: a test model for
merging biological intelligence into mechanical system. In: IEEE 7th Joint Interna-
tional Information Technology and Artificial Intelligence Conference (ITAIC-2014),
pp. 186–190 (2014)
9. Oda, T., Obukata, R., Ikeda, M., Barolli, L., Takizawa, M.: Design and imple-
mentation of a simulation system based on deep Q-network for mobile actor node
control in wireless sensor and actor networks. In: The 31th IEEE International Con-
ference on Advanced Information Networking and Applications Workshops (IEEE
WAINA-2017) (2017)
10. Oda, T., Elmazi, D., Cuka, M., Kulla, E., Ikeda, M., Barolli, L.: Performance evalu-
ation of a deep Q-network based simulation system for actor node mobility control
in wireless sensor and actor networks considering three-dimensional environment.
In: The 9th International Conference on Intelligent Networking and Collaborative
Systems (INCoS-2017), pp. 41–52 (2017)
11. Acampora, G., Cook, D., Rashidi, P., Vasilakos, A.V.: A survey on ambient intel-
ligence in health care. Proc. IEEE 101(12), 2470–2494 (2013)
12. Aarts, E., Wichert, R.: Ambient Intelligence. Technology Guide, pp. 244–249 (2009)
13. Aarts, E., de Ruyter, B.: New research perspectives on ambient intelligence. J.
Ambient Intell. Smart Environ. 1(1), 5–14 (2009)
14. Vasilakos, A., Pedrycz, W.: Ambient Intelligence, Wireless Networking, and Ubiq-
uitous Computing. Artech House, Norwood (2006)
15. Sadri, F.: Ambient intelligence: a survey. ACM Comput. Surv. 43(4), 36:1–36:66
(2011)
370 T. Oda et al.
16. Hagras, H., Callaghan, V., Colley, M., Clarke, G., Pounds-Cornish, A., Duman,
H.: Creating an ambient-intelligence environment using embedded agents. IEEE
Intell. Syst. 19(6), 12–20 (2004)
17. Ramos, C., Augusto, J.C., Shapiro, D.: Ambient intelligence-the next step for arti-
ficial intelligence. IEEE Intell. Syst. 23(2), 15–18 (2008)
18. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G.,
Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie,
C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S.,
Hassabis, D.: Human-level control through deep reinforcement learning. Nature
518, 529–533 (2015)
19. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D.,
Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv:1312.5602v1,
pp. 1–9 (2013)
20. Oda, T., Kulla, E., Cuka, M., Elmazi, D., Ikeda, M., Barolli, L.: Performance
evaluation of a deep Q-network based simulation system for actor node mobility
control in wireless sensor and actor networks considering different distributions of
events. In: The 11th International Conference on Innovative Mobile and Internet
Services in Ubiquitous Computing (IMIS-2017), pp. 36–49 (2017)
21. Oda, T., Kulla, E., Katayama, K., Ikeda, M., Barolli, L.: A deep Q-network based
simulation system for actor node mobility control in WSANs considering three-
dimensional environment: a comparison study for normal and uniform distribu-
tions. In: The 12th International Conference on Complex, Intelligent, and Software
Intensive Systems, pp. 842–852 (2018)
22. Lei, T., Ming, L.: A robot exploration strategy based on Q-learning network. In:
IEEE International Conference on Real-Time Computing and Robotics (RCAR-
2016), pp. 57–62 (2016)
23. Riedmiller, M.: Neural fitted Q iteration - first experiences with a data efficient neu-
ral reinforcement learning method. In: The 16th European Conference on Machine
Learning (ECML-2005), vol. 3720 of the series Lecture Notes in Computer Science,
pp. 317–328 (2005)
24. Lange, S., Riedmiller, M.: Deep auto-encoder neural networks in reinforcement
learning. In: The 2010 International Joint Conference on Neural Networks (IJCNN-
2010), pp. 1–8 (2010)
25. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially
observable stochastic domains. Artif. Intell. 101(1–2), 99–134 (1998)
26. Lin, L.J.: Reinforcement learning for robots using neural networks. Technical
report, DTIC Document (1993)
27. The Rust Programming Language (2007). https://www.rust-lang.org/
28. GitHub - rust-lang/rust: a safe, concurrent, practical language (2007). https://
github.com/rust-lang/
29. ‘rust’ tag wiki - Stack Overflow (2007). http://stackoverflow.com/tags/rust/info/
30. Ohno, Y.: Color rendering and luminous efficacy of white LED spectra. In: The
4th International Conference on Solid State Lighting, pp. 88–98 (2004)
31. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward
neural networks. In: The 13th International Conference on Artificial Intelligence
and Statistics (AISTATS-2010), pp. 249–256 (2010)
32. Glorot, X., Bordes, A., Bengio, Y.: Deep sparse rectifier neural networks. In: The
14th International Conference on Artificial Intelligence and Statistics (AISTATS-
2011), pp. 315–323 (2011)
An Efficient Scheduling of User
Appliances Using Multi Objective
Optimization in Smart Grid
1 Introduction
Energy is one of the most important resource, and energy demand is grow-
ing every day. Service providers are facing many problems to fulfill the energy
demand in residential building and industrial sectors. There are two ways to
solve this problem.
1. Produce additional energy and find new resources to produce energy
2. Excellent usage of existing resources
The first approach is costly and time consuming, as compared to the sec-
ond approach that is more efficient and inexpensive. Information technology
evolved and many schemes are introduced for energy consumption optimization.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 371–384, 2019.
https://doi.org/10.1007/978-3-030-15035-8_35
372 H. M. Faisal et al.
2 Related Work
Mixed integer linear programming (MILP), method is used in [1], to mini-
mize the total electricity bill paid by consumers is the main purpose of this
method. Authors worked on balance load management however, the user com-
fort is neglected in MILP. It shows an exchange between conventional systems
and today’s renewable energy sources (RES). A two-way communication between
utility and consumer through SM, a lot of energy wastage problem covered by
saving this 10–30%. In SG, the big challenge for researchers is cost minimization.
The genetic algorithm (GA) technique used in the paper [2], with the integration
of RES and stored energy a low level of cost minimization achieved. In particular
time changes when electricity price and user demand are higher than the stored
energy is helpful. Authors neglected the deployment and maintenance cost of
storage devices and RES in this technique. One of the big problem is to balance
the load in commercial and residential areas (Table 2).
374 H. M. Faisal et al.
technique is cost minimization. Customers can save maximum energy cost using
MINLP algorithm. Scheduling the load is the core objective of load management
during high demand to low demand time. Evolutionary algorithms are used for
load shifting [15]. All service sides have data sets, where scheduling problem have
been managed to solve the efficiency problem, the industry faced more problem
because of big power consumption appliances. Due to high load users need to use
the energy more intelligently in both residential and commercial sector. Authors
in [15] proposed an algorithm for load management. The main goal of this paper
reduces the electricity cost. All service sides have data sets, where the scheduling
problem have been managed to solve the efficiency problem, the industry faced
more problem because of big power consumption appliances. Customers require
high load in efficient way and more intelligently in the residential and commer-
cial area. Authors proposed a model for cost minimization in [17]. Using an
intelligent decision system (IDSS), minimum cost and minimum PAR problems
were solved. IDSS provides better result communication between the user and
utility. Authors discussed the EDE algorithm in [18]. An updated version of ED
was used. Authors used five trial vectors instead of one. Using three different
random vectors, a new population was created. The mutant and trial vectors
were generated by the fitness function. The authors proposed [19] evolutionary
accretive comfort algorithm (EASA) which is comprised of four postulations.
These postulations are defined according to the time and device bases priori-
ties. Based on the input values EASA generates optimal energy solution pattern
which satisfies the user budget. The author defines three different user budgets
to find the absolute solutions. Ma et al. [20] defines discomfort function for two
different type of gadgets. First category is flexible starting time and the other
is flexible power devices. Authors in [20] considered a multi-objective function
for user comfort and cost parameters. The proposed bat algorithm in [21] can
be applied to obtain the optimum result. By applying this algorithm energy
consumption can be reduced which is simply a non-linear optimization prob-
lem. The important goal of bat algorithm is to decrease the power usage and
increase consumer comfort standards in the residential area. In system models,
mathematical problems and scheduling, the bat algorithm used to solve these
problems.
3 Problem Statement
Total energy consumption, minimum electricity bill, minimum PAR and maxi-
mum user comfort are the most important problems in SG. DSM schemes are
used to overcome the aforementioned problems in the SG. In [1–4] purposed
meta heuristic techniques to solved the energy consumption problem according
to electricity bill and user comfort. Some authors used mathematical solutions
to solve these issues. Some authors [4–6] proposed RES with the integration of
ESS to tackle the energy demand problems. The main aim of all techniques is to
reduce the electricity bill with maximum user comfort. In some hours, when user
demand is higher than the total electricity production. It creates a peak load in
An Efficient Scheduling of User Appliances 377
the smart home so HEMS is faced difficulty to maintain the balance. To avoid
such problems, proposed algorithm, which shifts the home appliances from on
peak hours to off peak hours, and thus achieve the minimum cost and maximum
user comfort.
4 System Model
In SG, DSM controller is used for managing the user demand according to the
consumer demand. DSM provides more reliability and proficiency in user tasks.
Different meta heuristic techniques are implemented in DSM to control the user
appliances. DSM techniques manages the load to consume the electricity at off-
peak hours instead on peak hours. Initially, considering single home with 15
appliances. These appliances are categorized into two main categories: schedu-
lable and non-schedulable appliances. First category is further categorized into
two sub types: interruptible and non-interruptible. Non-interruptible appliances
are those appliances that cannot be shifted when they are running and cannot be
switched on as per the user’s requirements. Where interruptible appliances are
the those appliances which can be allocated to various time intervals. Every single
home contains SM. SM decides the operation time of every appliance according
to the power rating. These 15 appliances are taken from [14] details of appli-
ances are shown in Table 3. SM provides the two-way communication between
consumers and service provider. Different price signals (CPP, ToU, RTP) are
used to find the electricity bills. The service provider provides the electricity
price signal. ToU used as pricing unit to calculate the electricity cost. The main
objective of our study is to minimize electricity consumption in order to reduce
the cost and PAR, however, the tradeoff will occur between cost and user com-
fort. Equation 1 is used to calculate the PAR. Cost value is calculated by the
Eq. 2. Total load formula given in Eq. 3. Main architecture of system model is
shown in Fig. 1.
max(loads )
P AR = (1)
avg(loads )
24
App
hour
Cost = (ERate ∗ PRate ) (2)
t=1
app
Load = PRate ∗ App (3)
378 H. M. Faisal et al.
5 Proposed Scheme
the energy consumption problem and reduce the cost. The main objective of
EWO algorithm is to reduce the electricity bill and shift the appliances into off
peak hours. SALP algorithm provides the facility to manage the appliances using
different price signals. Different electricity price signals are discussed to define
the cost of electricity for a complete day. In our scheme, we consider RTP tariff.
The RTP is updated for every one hour. Two-way communication requires to
interact with the user for RTP.
5.1 EWO
• Each earthworm have the capacity to reproduce off springs and every earth-
worm individual have two kinds of reproduction,
• Every child of earthworm contains all the genetic factor of parents,
• Singular earthworm is moved on next generation, and cannot be changed by
operators.
5.2 SALP
Salps belongs to the family of salpidae. Swarming behavior is one of the most
interesting behavior. In salps, population has two groups: 1. Leader 2. Followers.
The leader is at the front of the chain and rest of the salps are attached behind
it followers. Equation 4 is used to update the position of the leader.
1 Fj + c1 ((ubj − lbj )c2 + lbj ) c3 ≥ 0
xj = (4)
Fj − c1 ((ubj − lbj )c2 + lbj ) c3 < 0
Our proposed algorithm provides the facility for consumers to schedule the appli-
ances and reduce the cost while considering the user comfort. The population
size is 30. In our proposed algorithm, the step of reproduction contains two types
of reproductions: Reproduction 1 and Reproduction 2. In our research, we have
implemented a new updated population scheme for scheduling the home appli-
ances. The first step is the initialization of all parameters with the maximum
generation and constant value. A fitness function is defined for choosing the
best solution. After applying the fitness function, two types of reproductions are
applied. The main purpose of the updated population algorithm is to obtain the
maximum solution of appliances. For complexity, mutation and crossover steps
are taken in the proposed algorithm. The major contribution of the proposed
algorithm updates the population according to their fitness function.
380 H. M. Faisal et al.
proportional to user comfort. By applying the price signal user comfort value
of proposed scheme is low, as compared to EWA and SSO. The user comfort is
reduced by 76.36%, 75.36% and 86.16% by the EWA, SSO and proposed scheme
respectively (Fig. 6).
Fig. 3. Load
382 H. M. Faisal et al.
Fig. 4. PAR
7 Conclusion
In this paper, proposed algorithm is used for shifting the appliances. The schedul-
ing is based on real-time data of price signal. Proposed scheme results are better
as compared to EWO and SALP algorithm. It is clear that the algorithm intro-
duced works efficiently as compared to EWO and SALP with the parameters of
cost, PAR and user comfort. With the proposed algorithm derived from EWO
and SALP proposed algorithm achieved minimum cost and maximum waiting
time of our proposed scheme. In the future, we will integrate renewable energy
system and ESS with more then one SH in order to minimize cost and maximize
user comfort.
References
1. Bradac, Z., Kaczmarczyk, V., Fiedler, P.: Optimal scheduling of domestic appli-
ances via MILP. Energies 8(1), 217–232 (2014)
2. Mary, G.A., Rajarajeswari, R.: Smart grid cost optimization using genetic algo-
rithm. Int. J. Res. Eng. Technol. 3(07), 282–287 (2014)
3. Bharathi, C., Rekha, D., Vijayakumar, V.: Genetic algorithm based demand side
management for smart grid. Wirel. Pers. Commun. 93(2), 481–502 (2017)
4. Setlhaolo, D., Xia, X., Zhang, J.: Optimal scheduling of household appliances for
demand response. Electr. Power Syst. Res. 116, 24–28 (2014)
5. Samadi, P., Wong, V.W.S., Schober, R.: Load scheduling and power trading in
systems with high penetration of renewable energy resources. IEEE Trans. Smart
Grid 7(4), 1802–1812 (2016)
6. Ullah, I., Javaid, N., Khan, Z.A., Qasim, U., Khan, Z.A., Mehmood, S.A.: An
incentive-based optimal energy consumption scheduling algorithm for residential
users. Procedia Comput. Sci. 52, 851–857 (2015)
7. Cakmak, R., Altas, I.H.: Scheduling of domestic shiftable loads via cuckoo search
optimization algorithm. In: 2016 4th International Istanbul Smart Grid Congress
and Fair (ICSG), pp. 1–4. IEEE, April 2016
8. Ahmed, M.S., Mohamed, A., Khatib, T., Shareef, H., Homod, R.Z., Ali, J.A.: Real
time optimal schedule controller for home energy management system using new
binary backtracking search algorithm. Eng. Build. 138, 215–227 (2017)
9. Huang, Y., Wang, L., Guo, W., Kang, Q., Wu, Q.: Chance constrained optimization
in a home energy management system. IEEE Trans. Smart Grid (2016). https://
doi.org/10.1109/TSG.2016.2550031
10. Nguyen, H.K., Song, J.B., Han, Z.: Distributed demand side management with
energy storage in smart grid. IEEE Trans. Parallel Distrib. Syst. 26, 3346–3357
(2015)
11. Gao, X.Z., Govindasamy, V., Xu, H., Wang, X., Zenger, K.: Harmony search
method: Theory and applications. Comput. Intell. Neurosci. 2015, 39 (2015).
https://doi.org/10.1155/2015/258491
12. Mahmood, A., Baig, F., Alrajeh, N., Qasim, U., Khan, Z.A., Javaid, N.: An
enhanced system architecture for optimized demand side management smart grid.
Appl. Sci. 6, 122 (2016). https://doi.org/10.3390/app6050122
13. Zhang, J., Wu, Y., Guo, Y., Wang, B., Wang, H., Liu, H.: A hybrid harmony
search algorithm with differential evolution for day-ahead scheduling problem of a
microgrid with consideration of power flow constraints. Appl. Energy 183, 791–804
(2016)
384 H. M. Faisal et al.
14. Shirazi, E., Zakariazadeh, A., Jadid, S.: Optimal joint scheduling of electrical and
thermal appliances in a smart home environment. Energy Convers. Manag. 106,
181–193 (2015)
15. Logenthiran, T., Srinivasan, D., Shun, T.Z.: Demand side management in smart
grid using heuristic optimization. IEEE Trans. Smart Grid 3(3), 1244–1252 (2012)
16. Yuce, B., Rezgui, Y., Mourshed, M.: ANN-GA smart appliance scheduling for
optimised energy management in the domestic sector. Energy Build. 111, 311–325
(2016)
17. Ogwumike, C., Short, M., Abugchem, F.: Heuristic optimization of consumer elec-
tricity costs using a generic cost model. Energies 9, 6 (2016)
18. Islam, S.M., Das, S., Ghosh, S., Roy, S., Suganthan, P.N.: An adaptive differential
evolution algorithm with novel mutation and crossover strategies for global numer-
ical optimization. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 42, 482–500
(2012)
19. Khan, A., Javaid, N., Khan, M.I.: Time and device based priority induced comfort
management in smart home within the consumer budget limitation
20. Ma, K., Yao, T., Yang, J., Guan, X.: Residential power scheduling for demand
response in smart grid. Int. J. Electr. Power Energy Syst. 78, 320–325 (2016)
21. Mohan, N., Sivaraj, R., Pyira, R.D.: A comprehensive reviews of bat algorithm and
its application to various optimization problems. Asian J. Res. Soc. Sci. Humanti.
6, 676–690 (2016)
Pro Utility Pro Consumer Comfort
Demand Side Management in Smart Grid
1 Introduction
The smart grid is more powerful than the traditional grid system because of bidi-
rectional communication between the source and end users which makes it more
scalable and efficient [1]. The Demand Response (DR) helps in the interaction
between the power source and consumers to provide more efficient and reliable
system [2]. The main role of DR is to reduce the burden on the power source
by motivating the consumers to change their energy usage behavior according
to price tariff provided by a utility. It has been analyzed that more than 40% of
energy is consumed in a residential area around the world [3].
The international energy outlook in 2013 assumes that by 2040, it will rise to
more than 56% [4]. To tackle all these challenges, smart grid emerged with the
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 385–397, 2019.
https://doi.org/10.1007/978-3-030-15035-8_36
386 W. Ahmad et al.
2 Related Work
single objective optimization. This system schedule the smart devices of an intel-
ligent home. Furthermore, the objective was to stable the load by a mixed integer
linear programming to reduce the cost and PAR. The simulation conducted in
the Czech Republic by taking real-time pricing.
In [14,17], the article focus was to reduce the total cost and reduce the
electricity bill of an individual. Simulations in [18], performed on real-time data
received from Energy and Climate Center UK by taking real time tariffs with
multi-objective mixed integer programming. The high demand determination
scheme is proposed in [15], for residential area using Real Time Pricing (RTP)
for multiple cases. The finite smart appliances are taken with quasi-random
methods and the recursive equation is formed to check the high load demand.
Intelligent HEMS model is presented in [19]. In this model, oU pricing tariff
is used to reduce the total cost without affecting the consumer satisfaction.
In this concern [20], a cost-effective greedy iterative distributed model is used
for multiple consumers to overcome the cost problem in a grid. The simulation
outcomes expose that electricity cost is degraded and PAR in decreased. The
distributed generation model is proposed in [21], which effectively reduced the
cost and waiting time by allowing bidirectional electricity flow. Results show
that for all cases, the 32.73% of energy cost is saved. The user discomfort is
discussed in [22]. In this research paper, the demand response issue is developed
in HEMS with a support of the Markov decision model. The DSM controller is
employed with Genetic Algorithm (GA), Binary Particle Swarm Optimization
(BPSO) and Colony Optimization Algorithm (COA) in [23]. In this scheme, GA
outperforms the rest of the schemes. The cost minimization and waiting time is
directly related to user comfort as discussed in [24].
3 Problem Statement
Traditional grid system has many flaws and limitations. From the last decade,
smart grid attracts the attention of the researcher towards its benefits and advan-
tages. The smart grid contains two-way communication between consumer and
power source. This absolute goal of the smart grid to generate enhancement
within supply, to organize and regulate the whole process of electrical energy
[25]. To meet the requirement of consumers for energy demand by utilizing dif-
ferent techniques identified by DSM. As we know, a single home has too many
electrical appliances which are running time to time with the different type of
power ratings. With the growth in population, power need in the residential side
is raising gradually [26]. By means of employing DSM, a person can help to
eliminate the price as well as handle the unwanted using electrical power by way
of arrangement appliances properly. Through DSM, scheduling of smart devices
is done by knowing the peak hour and estimating the price through the present
hour. These kinds of switching lesson the price but load the utility and therefore
associated with general shortage of energy.
388 W. Ahmad et al.
4 Proposed Solution
The main focus of our work is to reduce the peak load and total cost which is
calculated in Eq. 1. The RTP and ToU 24-h onward day pricing is used in the
simulation. The central challenging part is load shifting from On-Peak hours to
Off-peak hours in order to tackle to load balance and reduce cost. The load shift-
ing is done based on the fitness criterion which is presented in Eq. 2. Now in order
to determine the total load, the Eq. 3 is employed, where alpha represents the
status of appliances in 0’s and 1’s. Table 1 shows the appliances and their power
rating according to their classifications of interruptible and non-interruptible.
Pro Utility Pro Consumer 389
24
App
hour
Cost = (ERate ∗ PRate ) (1)
hour=1
⎧ ieN p
⎨lod ≥ mean(LU s hour
od ), ERate ≤ mean(ERate )
Np ieN p
Ff = min l > (std(Lod )) ∧ lod < mean(LU
U s s
od )
(2)
⎩ od hour
ERate > mean(ERate )
App
Lod = PRate ∗ App (3)
Our main objective is not only lessened the Cost as defined in Eq. 5 but also
decrement in PAR by applying Eq. 6. Along with cost and PAR reduction, we
also have objective to achieve efficient load shifting in DSM as assessed in Eq. 7.
Ojb1 = min(Cost) (4)
Ojb2 = min(P AR) (5)
Ojb3 = min(Load) (6)
To achieve this goal, we have to split a day into 24-h time slots. Furthermore,
these time slots are split into two parts (high price hours and low price hours).
Now the main challenge is to transfer the high load to low load hours according
to their power ratings classification. This approach will effectively decrease the
cost and PAR due to load balance. For calculation of PAR, we have used Eq. 8.
max(LSod )
P AR = (7)
Average(LSod )
5 System Model
In our proposed work, simulation is performed on a smart home which consists
of fifteen different appliances that need to be scheduled. Smart appliance Length
of Operation (LoT) and power information is collected from end users. The pro-
posed system model is divided into three sections such as DSM, Supply Side
Management (SSM), and Communication unit. SSM have all the information
related to energy generation and pricing tariff. Energy Management Controller
(EMC) is used by DSM for appliance scheduler. The objective of handling the
load is to bound the consumers from excessive usage of electricity to lower the
burden on utility. The communication unit is responsible for the transfer of
the pricing scheme and dynamic demand-response data between DSM and SSM.
Home Area Network (HAN) is used in a smart home for communication between
the smart appliances and EMC through Wi-Fi and Zig-Bee protocols. Further-
more, there are three classifications of smart devices according to the schedule:
the Interruptible, Un-interruptible, and base load. In our paper, three meta-
heuristic scheme are selected, including FA, BFA and RCES. These schemes are
adopted to manage the smart appliance in HEMS. The central idea of this sys-
tem is to decrease the consumption throughout the peak time to diminish the
cost and PAR. Simulations is performed on actual presumed smart appliances
that are usually split by a few distinct types like: Interruptible, Non-interruptible
and base load (Fig. 1) (Table 2).
390 W. Ahmad et al.
6 Meta-heuristic Algorithms
Several optimization methods have already been deployed to handle the smart
grid problems. In this regard, several effective heuristic techniques have already
been utilized to produce a strong strategy to tackle energy and smart grid prob-
lems. In that study, we have applied two meta-heuristic systems, BFA and FA.
6.1 BFA
Nature neglects those animals which have bad foraging methods and support
those who have surviving strategies. Initially, BFA algorithm was developed
by [27]. After many productions, poor versions usually are changed or even
Pro Utility Pro Consumer 391
6.2 FFA
FFA merge as an obstacle solver [28]. This algorithm helps in giving optimized
solutions to many different optimizations obstacles. Firefly is known as multi-
handler due to its multi-model feature. In our research, we have adopted a new
scheme to tackle the scheduling index.
6.3 RCES
In our research, we have implemented a new Random Cell Elimination Scheme
(RCES). In RCES, random data is generated and the distance of the closest cell
is evaluated. The population is assessed by the number of steps in the defined
position and fitness function is defined to cut back the overall cost. The pro-
duction of the Cell is evaluated and replaced with the fittest position. The new
generation of cell is produced in next step with random population. This new
generation eliminates the old cells based on objective and fitness evaluation. The
main purpose of RCES is o obtain the optimum solution of appliance scheduling
in smart home. However, optimized home scheduling task is a difficult goal to
achieve. Simulation results demonstrate that our system successfully lower the
price and PAR as trade-off between price and delay time, so in order to achieve
one factor, other might be sacrificed.
1. Initialization of the population
2. Fitness Evaluation of each Cell
3. Reproduction
4. Replace previous place with good one.
The main objective of our approach is to reduce the PAR to lower the burden
on the utility and reduce waiting time. PAR explain the power consumption pat-
tern of the user load and effects the power grids. In Figs. 6 and 7, unscheduled and
scheduled PAR is illustrated. The simulation results shows that presented app-
roach RCES reduce considerably amount of PAR as compared to BFA and FA,
which shows that proposed approach is better than other two heuristic schemes.
However, PAR is related to burden on power grids. Simulation results reveal all
the heuristic schemes minimize the PAR. However our proposed heuristic opti-
mization scheme reduce PAR to 80% and beats the other schemes. So we can say
that our new scheme is pro utility and pro user comfort achievement in smart
grid.
Our New proposed scheme also shows the effectiveness in reducing total cost
as compared to unscheduled scenario. All heuristic algorithm shows effective-
ness in reducing the cost. Reducing cost and PAR using heuristic algorithms is
challenging objective because of trade off between cost and delay time. Figures 8
and 9, the results represent the total energy cost for BFA, FA and RCES using
ToU and RTP pricing tariffs.
Figures 10 and 11, present the waiting time which is mainly concern to user
comfort. User comfort is achieved when a consumer faces less waiting time in
operation of electrical appliances. The results shows that our approach is a pro
394 W. Ahmad et al.
user comfort because of increase in the user comfort by reducing the waiting
time to 50% which is a major contribution of our work.
In order to analyze the model, the confidence level is measured in term of PAR
and waiting time. Simulation results indicates less gap even in random values.
BFA is an old optimization technique which mainly focuses on production and
Pro Utility Pro Consumer 395
living of bacteria, while Firefly has known as problem solver because they work
on flashing behavior and distance of the fittest between them. The results reveal
that RCES is efficiently reduce the cost and PAR while scheduling the appliances
towards On-Peak to Off-Peak hour without any stress on the Power Grid.
396 W. Ahmad et al.
8 Conclusion
DSM is used for the control and management of electrical appliances used in
a smart home in order to balance the load demand. In most cases the energy
load is shift to new renewable sources to ease the user but it is not sufficient.
To tackle the above mention problem an optimization technique is adopted. The
new scheme RCES has been proposed which efficiently reduce PAR and user
discomfort. The experiments results clearly revealed that our propose scheme is
better than BFA and FA. The high load is shifted to low peak hours in order
to balance the load and reduce the stress on utility. It is observed that there is
always a tradeoff between cost, PAR and user comfort. The results show that
fitness function improve the performance. The experiment is conducted on the
day ahead ToU and RTP pricing schemes which originally examine the behavior
of optimization schemes. To check the performance of the ToU and RTP pricing
schemes, price forecasting is an essential feature.
References
1. Ipakchi, A., Albuyeh, F.: Grid of the future. IEEE Power Energy Mag. 7(2), 52–62
(2009)
2. Medina, J., Muller, N., Roytelman, I.: Demand response and distribution grid
operations: opportunities and challenges. IEEE Trans. Smart Grid 1(2), 193–198
(2010)
3. Gul, M.S., Patidar, S.: Understanding the energy consumption and occupancy of
a multi-purpose academic building. Energy Build. 87, 155–165 (2015)
4. Khalid, A., et al.: Towards dynamic coordination among home appliances using
multi-objective energy optimization for demand side management in smart build-
ings. IEEE Access 6, 19509–19529 (2018)
5. Gellings, C.W.: The concept of demand-side management for electric utilities. Proc.
IEEE 73(10), 1468–1470 (1985)
6. Broehl, J.H., Huss, W.R., Skelton, J.C., Sullivan, R.W., Limaye, D.R., Sloane,
B.D., Davis, T.D., Blevins, R.P.: Demand-side management, vol. 1. Overview of
key issues. Final report. No. EPRI-EA/EM-3597-Vol. 1. Battelle Columbus Labs.,
OH (USA); Synergic Resources Corp., Bala-Cynwyd, PA (USA) (1984)
7. Davito, B., Tai, H., Uhlaner, R.: The smart grid and the promise of demand-side
management. McKinsey Smart Grid 3, 8–44 (2010)
8. Gelazanskas, L., Gamage, K.A.: Demand side management in smart grid: a review
and proposals for future direction. Sustain. Cities Soc. 11, 22–30 (2014)
9. Logenthiran, T., Srinivasan, D., Shun, T.Z.: Demand side management in smart
grid using heuristic optimization. IEEE Trans. Smart Grid 3(3), 1244–1252 (2012)
10. Shirazi, E., Jadid, S.: Optimal residential appliance scheduling under dynamic
pricing scheme via HEMDAS. Energy Build. 93, 40–49 (2015)
11. Matayoshi, H., Uehara, T., Ludin, G.A., Sharma, A., Datta, M., Senjyu, T.: Unin-
terruptible smart house equipped with a single-phase dq-transformation system.
J. Renew. Sustain. Energy 8(2), 025101 (2016)
12. Mesari, P., Krajcar, S.: Home demand side management integrated with electric
vehicles and renewable energy sources. Energy Build. 108, 1–9 (2015)
Pro Utility Pro Consumer 397
13. Adika, C.O., Wang, L.: Smart charging and appliance scheduling approaches to
demand side management. Int. J. Electr. Power Energy Syst. 57, 232–240 (2014)
14. Gao, B., Zhang, W., Tang, Y., Mingjin, H., Zhu, M., Zhan, H.: Game-theoretic
energy management for residential users with dischargeable plug-in electric vehi-
cles. Energies 7(11), 7499–7518 (2014)
15. Vardakas, J.S., Zorba, N., Verikoukis, C.V.: Power demand control scenarios for
smart grid applications with finite number of appliances. Appl. Energy 162, 83–98
(2016)
16. Zhang, Y., Zeng, P., Li, S., Zang, C., Li, H.: A novel multiobjective optimization
algorithm for home energy management system in smart grid. Math. Probl. Eng.
2015, 19 (2015)
17. Jovanovic, R., Bousselham, A., Bayram, I.S.: Residential demand response schedul-
ing with consideration of consumer preferences. Appl. Sci. 6(1), 16 (2016)
18. Qdr, Q.: Benefits of demand response in electricity markets and recommendations
for achieving them. US Department of Energy, Washington, DC, USA, Technical
report (2006)
19. Abushnaf, J., Rassau, A., Grnisiewicz, W.: Impact on electricity use of introducing
time of use pricing to a multiuser home energy management system. Int. Trans.
Electr. Energy Syst. 26(5), 993–1005 (2016)
20. Chavali, P., Yang, P., Nehorai, A.: A distributed algorithm of appliance schedul-
ing for home energy management system. IEEE Trans. Smart Grid 5(1), 282–290
(2014)
21. Elkazaz, M.H., Hoballah, A., Azmy, A.M.: Artificial intelligent based optimization
of automated home energy management systems. Int. Trans. Electr. Energy Syst.
26(9), 2038–2056 (2016)
22. Wen, Z., O’Neill, D., Maei, H.: Optimal demand response using device-based rein-
forcement learning. IEEE Trans. Smart Grid 6(5), 2312–2324 (2015)
23. Rahim, S., Javaid, N., Ahmad, A., Khan, S.A., Khan, Z.A., Alrajeh, N., Qasim, U.:
Exploiting heuristic algorithms to efficiently utilize energy management controllers
with renewable energy sources. Energy Build. 129, 452–470 (2016)
24. Rasheed, M.B., Awais, M., Javaid, N., Iqbal, Z., Khurshid, A., Chaudhry, F.A.,
Ilahi, F.: An energy efficient residential load management system for multi-class
appliances in smart homes. In: 2015 18th International Conference on Network-
Based Information Systems (NBiS), pp. 53–57. IEEE (2015)
25. California Energy Commission. Energy Efficiency and Conservation: Trends and
Policy Issues. No. 100-03. Report, May 2003
26. Jacob, L., Jeyakrishanan, V., Sengottuvelan, P.: Resource scheduling in cloud using
bacterial foraging optimization algorithm. Int. J. Comput. Appl. 92(1), 14–20
(2014)
27. Passino, K.M.: Biomimicry of bacterial foraging for distributed optimization and
control. IEEE Control Syst. 22(3), 52–67 (2002)
28. Fister, Jr., I., Yang, X.-S., Fister, I., Brest, J., Fister, D.: A brief review of nature-
inspired algorithms for optimization. arXiv preprint arXiv:1307.4186 (2013)
Efficient Scheduling of Smart Home
Appliances for Energy Management
by Cost and PAR Optimization
Algorithm in Smart Grid
1 Introduction
The traditional Grid (TG) has insufficient capabilities to solve electricity grid chal-
lenges: security of transmission line, bi-directional communication scalability and
robustness against any fault [1]. Therefore, advanced architecture of the TG is
strongly wished to overcome these challenges in an efficient way. Smart Grid (SG)
integrates various communications information to TG. SG allow consumer to con-
trol electricity consumption through bi-directional communication between con-
sumers and source via Smart Meter (SM) in Advanced Metering Infrastructure
(AMI) [2]. Demand Side Management (DSM) is the main section of SG to cre-
ate balance for both demand and supply sides [3]. DSM has two main features:
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 398–411, 2019.
https://doi.org/10.1007/978-3-030-15035-8_37
Efficient Scheduling of Smart Home Appliances 399
Demand Response (DR) action taken by consumer on dynamic pricing scheme and
load management to schedule the electricity consumption in an efficient way. Elec-
tricity consumption is optimize to reduce electricity cost by DR Program in [4] i.e.
different electricity consumption patter dynamic electricity price.
The main concern of SG is to reduce the PAR, electricity bills and maximize
the consumer comfort. Electricity bills and PAR are reduced at demand side by
efficient management of electricity consumption pattern. Utility provide differ-
ent pricing scheme: Real Time Price (RTP), Critical Peak Price (CPP), Inclined
Block Rate (IBR) etc. Electricity can be managed for the home appliances for
reduction of cost and Peak to Average Ratio (PAR) by pricing schemes. Cur-
rently many DSM mechanisms deployed to overcome previously mention chal-
lenges. In past research work, many of non-heuristic technique are presented to
optimize the home appliances [5], the researcher used Integer Linear Program-
ming (ILP) for reducing load at low price time slot to increase consumer comfort
(CC). However, this technique is not consider for complex energy consumption
pattern.
Scheduling of residential appliances in architecture of SG is provided by Home
Energy Management System (HEMS). HEMS is an intelligent system that opti-
mizes the load consumption pattern in peak hours. In previous work of HEMS,
authors categories the load management as real-time and predictable based. In
[6], predictable approach of load management is deployed. However, this app-
roach is costly and complex to produce uncertainty of solution. While real-time
optimization is employed to overcome the uncertainty issues that is deployable
for large scale of area [7].
In [8], real-time approach of load management is deployed for peak hours.
The exclusively management of energy utilization pattern in residential area is
feasible for optimization of different appliances at home. However, CC is not
satisfied. In this paper, we deployed the meta-heuristic algorithm for real-time
environment. In our work, Strawberry Algorithm (SA) and Firefly Algorithm
(FA) are simulated for their results to compare with our proposed scheme Runner
Updation Optimization Algorithm (RUOA). We developed a new scheme to get
optimal solution for reduction of total electricity cost and PAR as compared
with SA and FA scheme, while CC is sacrificed in our scheme. However, there
is trade-off exist for CC to cost and PAR. Moreover, many spaces exist to fill
the area of load management and more researches are going on to overcome the
energy crises.
The rest of the work is organized as follows: In Sect. 2, we present the related
work for scheduling of appliances in according with different technique. Section 3
explains the problem statement to tackle the issues. In Sect. 4, we discussed
our proposed scheme in detail along with pricing scheme that has been used.
Scheduling techniques are presented in Sect. 5 to work out our proposed scheme.
Simulation results are discussed in Sect. 6. At the end paper conclusions are
drawn in Sect. 7.
400 S. M. Shuja et al.
2 Related Work
In HEMS lot of research has been presented by many authors for optimization
of residential appliances’ scheduling. Researchers trying to induce the efficient
scheduling of electricity consumption pattern by the appliances placed at any
home. Different authors have main objective to consumed electricity in reliable
and efficient way by reducing electricity cost and PAR or increasing CC. Further,
some related research works are discussed below and summarized in Table 1.
Authors highlighted the multi residential electricity load scheduling problem
in [9]. According to author, in previous research only single residential area is
considered, to keep this thing in mind they want to maximize the user comfort for
large residential area. They proposed electricity load scheduling algorithm named
as PL-Generalized Benders Algorithm. The aim of their proposed algorithm is
to schedule the load of appliances and maximize the user comfort. However, it
is not defined how will they schedule load when appliances are near optimal to
each other.
Discussion about the trend by authors in [10], to schedule the electrical load
from on peak hours to off peak hours. The consumer needs scheduling in an
online manner that they can easily find out the prices of electricity and schedule
the load of their appliances according to that manner. To make electricity users
aware about their load scheduling of appliances online Load Scheduling Learning
(LSL) algorithm is proposed in [10]. The aim of their algorithm is to reduce PAR.
However, cost of electricity and performance of proposed algorithm is not defined
by them.
Community based cooperative energy scheme is proposed in [11]. The aim of
proposed scheme is to discuss the electricity cost consumption of the user and
smart grid. In their work, they want to minimize the cost and PAR, to fulfill
their aim they considered community between SG to consumer and consumer to
SG. To evaluate the results of their proposed C2C (community to community)
scheme, MATLAB is used. The cost is minimized because of reducing PAR dur-
ing on peak hours. However, User comfort and security issues are compromised.
In [12], the HEMS system is proposed to schedule the operation of electric
appliances. Authors perceive about Quality of Experience in their work because
they want to check the effects of the proposed system. For this purpose they
consider the profiles of the users and also proposed allocation algorithm. By
applying the proposed system and algorithm electricity cost is minimized. On
the other hand, they have not consider PAR and CC in their proposed work.
The Binary Particle Swarm Optimization (BPSO), Genetic Algorithm (GA)
and cuckoo search meta-heuristic algorithm is used in [13]. In their paper, the aim
of authors is to schedule the load of homes appliances and convert them in smart
homes. They also want to reduce the peak load and electricity bill reduction.
To schedule the appliances load of the homes they also used renewable energy
resources in their work. The electricity bill is minimized. However, the electricity
bill can be reduced further.
The multiple problems such as reducing electricity bill cost and peak reduc-
tion is considered in [14]. They want to solve above mentioned problems, to
Efficient Scheduling of Smart Home Appliances 401
3 Problem Statement
The proposed system model is shown in Fig. 1. Our system model is working for
scheduling SH appliances. Cost of electricity and PAR is increasing periodically
so there is need of well adaptable system, which can minimize average electricity
price and PAR. A HEMS is used for optimization of home appliances to regulate
the schedule of their operational time. HEMS ease the consumer with respect
of reduction in electricity dissipation and electricity cost. For optimization the
appliances are classified into three different categories; Shiftable appliances, Non-
Shiftable appliances and fixed appliances. Shiftable are those appliances whose
operational time can be shifted from peak hour time slots to non-peak hour time
402 S. M. Shuja et al.
4 Scheduling Techniques
Many of the mathematical techniques are presented for scheduling of residen-
tial appliances; MILP, DP, MINLP etc. However, due to slow computation for
large number of appliances consumer demand is not satisfied. Therefore, meta-
heuristic algorithms are used for efficient optimization of home appliances to
reduce the electricity price and manage the load between peak and non-peak
hour. In this paper, we choose some meta- heuristic algorithm SBA and FA to
propose our algorithm RUOA for optimization of energy consumption pattern
of appliances. Algorithms are described further in below subsection.
results of proposed algorithm RUOA. In our work, we compared the result with
existing meta-heuristic SBA and FA algorithm. CPP and RTP pricing scheme
are used for our main objective of cost minimization, PAR reduction and load
management in simulations. Figures 2 and 3 shows the hourly electricity cost
along with plots of pricing signal CPP and RTP.
The hourly load consumptions are shown in Fig. 4 for CPP and Fig. 5 for RTP
along with scheduled and unscheduled pattern of consumption. Our proposed
algorithm RUOA show that its peaks for hourly load graph are better than other
meta-heuristic algorithm SBA and FA for both CPP and RTP. We realize that pat-
tern of per hour load consumption of schedule is much better than unscheduled.
Our simulation result depict that proposed algorithm RUOA optimize the hourly
load consumption by shifting load from peak hours to non-peak hours.
The result of PAR value generated by SBA, FA and proposed scheme RUOA
for CPP and RTP are shown in Figs. 6 and 7. Both CPP and RTP signals in
figure verify that proposed algorithm outperformed than other algorithm. PAR
is reduced by 70.76% in SBA, 45.21% in FA and 75.1% in proposed RUOA
for CPP, while 65.45% in SBA, 45.2% in FA and 75.02% in proposed RUOA
for RTP with refer to unscheduled PAR. Simulations show that our proposed
scheme RUOA performed in a better way than other algorithm.
The hourly cost of electricity shown in Figs. 8 and 9 for CPP and RTP, which
simulated by SBA, FA and our proposed algorithm RUOA. Simulations depict
the result of our proposed scheme is good enough from other scheme, as schedule
peak of graph show the lower cost at higher price time slot as compared with
unscheduled peaks.
Total electricity cost value is estimated in Figs. 10 and 11 simulated by SBA,
FA and RUOA for CPP and RTP pricing scheme. Simulation shows that perfor-
mance of our proposed scheme is outclassing as compared with SBA and FA. In
Total Cost plots, the total electricity cost is reducing by 24.53% in SBA, 55.4%
in FA and 65.97% in proposed RUOA for CPP, while 19.78% in SBA, 15.21% in
FA and 35.19% in proposed RUOA with respect of unscheduled case. There is
trade-off between cost and waiting time, which mean that our proposed scheme
RUOA is reducing cost while sacrificing consumer waiting time.
Consumer waiting time for CPP and RTP pricing signal are presented in
Figs. 12 and 13. In our case, waiting time shows the time limit to turn ON the
home appliances. The consumer comfort level is calculated in term of consumer
waiting time, we have to minimize the waiting time for consumer to maximize the
comfort level and vice versa. The waiting time calculated for CPP is 3.8876 h by
SBA, 6.7775 h by FA and 4.9879 h by proposed RUOA, while for RTP is 3.8409 h
by SBA, 6.7775 h by FA and 4.5403 h by proposed RUOA. Simulation shows that
Efficient Scheduling of Smart Home Appliances 409
SBA well performed than other algorithm in consumer waiting time. However,
there is trade-off for our proposed RUOA to SBA and FA in waiting time and
total electricity price.
6 Conclusion
An electricity load management is concerned to evaluate at DSM. In this
paper, we develop the scheme for load management from existing meta-heuristic
410 S. M. Shuja et al.
technique SBA and FA with pricing scheme of CPP and RTP. We proposed an
RUOA algorithm to optimize the residential appliances based on their electric-
ity consumption pattern. We analyzed the performance of our simulation, which
shows that the performance of our proposed technique is better than other meta-
heuristic technique SBA and FA in term of total electricity price and PAR. On
the other hand SBA outperformed for consumer comfort as compared with our
proposed algorithm.
References
1. Gungor, V.C., Sahin, D., Kocak, T., Ergut, S., Buccella, C., Cecati, C., et al.:
Smart grid technologies: communication technologies and standards. IEEE Trans.
Ind. Inform. 7(4), 529–539 (2011)
2. Esther, B.P., Kumar, K.S.: A survey on residential demand side management archi-
tecture, approaches, optimization models and methods. Renew. Sustain. Energy
Rev. 59, 342–351 (2016)
3. Wu, Z., Tazvinga, H., Xia, X.: Demand side management of photovoltaic-battery
hybrid system. Appl. Energy 148, 294–304 (2015)
4. Strbac, G.: Demand side management: benefits and challenges. Energy Policy 36,
4419–4426 (2008). Vancouver
5. Zhu, Z., Jie, T., Lambotharan, S., Chin, W.H., Fan, Z.: An integer linear pro-
gramming based optimization for home demand-side management in smart grid.
In: 2012 IEEE PES Innovative Smart Grid Technologies (ISGT), pp. 1–5. IEEE
(2012)
6. Agnetis, A., de Pascale, G., Detti, P., Vicino, A.: Load scheduling for household
energy consumption optimization. Smart Grid IEEE Trans. 4, 2364–2373 (2013)
7. Colmenar-Santos, A., de Lober, L.N.T., Borge-Diez, D., Castro-Gil, M.: Solutions
to reduce energy consumption in the management of large buildings. Energy Build.
56, 66–77 (2013)
8. Marzband, M., Ghazimirsaeid, S.S., Uppal, H., Fernando, T.: A real-time evalu-
ation of energy management systems for smart hybrid home Microgrids. Electr.
Power Syst. Res. 143, 624–633 (2017)
9. Moon, S., Lee, J.-W.: Multi-residential demand response scheduling with multi-
class appliances in smart grid. IEEE Trans. Smart Grid 9(4), 2518–2528 (2018)
10. Bahrami, S., Wong, V.W.S., Huang, J.: An online learning algorithm for demand
response in smart grid. IEEE Trans. Smart Grid 9(5), 4712–4725 (2018)
11. Bera, S., Misra, S., Chatterjee, D.: C2C: community-based cooperative energy
consumption in smart grid. IEEE Trans. Smart Grid 9(5), 4262–4269 (2018)
12. Behrens, D., Schoormann, T., Knackstedt, R.: Developing an algorithm to consider
mutliple demand response objectives. Eng. Technol. Appl. Sci. Res. 8(1), 2621–2626
(2018)
13. Javaid, N., Ullah, I., Akbar, M., Iqbal, Z., Khan, F.A., Alrajeh, N., Alabed, M.S.:
An intelligent load management system with renewable energy integration for
smart homes. IEEE Access 5, 13587–13600 (2017)
14. Aslam, S., Iqbal, Z., Javaid, N., Khan, Z.A., Aurangzeb, K., Haider, S.I.: Towards
efficient energy management of smart buildings exploiting heuristic optimization
with real time and critical peak pricing schemes. Energies 10(12), 2065 (2017)
15. Khan, I.U., Ma, X., Taylor, C.J., Javaid, N., Gamage, K.: Heuristic algorithm
based dynamic scheduling model of home appliances in smart grid (2018)
Efficient Scheduling of Smart Home Appliances 411
16. do Prado, J.C., Qiao, W.: A stochastic decision-making model for an electricity
retailer with intermittent renewable energy and short-term demand response. IEEE
Trans. Smart Grid (2018)
17. Ma, K., Yao, T., Yang, J., Guan, X.: Residential power scheduling for demand
response in smart grid. Int. J. Electr. Power Energy Syst. 78, 320–325 (2016). 129,
452–470
18. Marzband, M., Yousefnejad, E., Sumper, A., Domnguez-Garca, J.L.: Real time
experimental implementation of optimum energy management system in stan-
dalone microgrid by using multi-layer ant colony optimization. Int. J. Electr. Power
Energy Syst. 75, 265–274 (2016)
19. Merrikh-Bayat, F.: A Numerical Optimization Algorithm Inspired by the Straw-
berry Plant. arXiv 2014, arXiv:1407.7399 (2014)
20. Yang, X.-S.: Firefly Algorithms for multimodal optimization. In: International
Symposium on Stochastic Algorithms. Springer, Berlin (2009)
Multiple S-Box Correlation Energy Analysis
Model Based on Particle Swarm Optimization
1 Introduction
In 2004, Eric Brier proposed a new type of cipher attack method—correlation energy
analysis [1]. This technology uses the Pearson correlation coefficient [2] p in statistics to
analyze the attack, and recover the key to achieve the attack effect. Compared with other
key recovery technologies, it has stronger attack and higher decryption efficiency.
However, in the traditional correlation energy analysis [3], the correlation between the S-
box and the energy curve is poor [4]. Simply attacking single S-box will turn the
remaining S-box into noise, which will affect the efficiency and accuracy of key cracking.
To enhance the correlation between the S-box and the energy curve [5], multiple S-boxes
can be considered as a whole, and be attacked in parallel, but the number of keys guessed
is too large in this situation, which will cause the key to be unrecoverable. For example,
when attacking four S-boxes simultaneously, the number of bits of the key guessed is
four times that of attacking single S-box, who increases exponentially [6].
With the spreading of artificial intelligence optimization algorithm [7], the advan-
tages of some intelligent optimization algorithms are further reflected. The artificial
intelligence algorithm provides convenience of solving the problem of excessive com-
putation in the process of attacking multiple S-boxes simultaneously with its high
computational efficiency and strong reliability. In this paper, a correlation energy anal-
ysis model based on particle swarm optimization (PSO) algorithm is proposed, according
to the characteristic of the encryption key. The DES is used as the experimental object to
carry out the CPA attack experiment. The experimental results show that Our model can
effectively improve the speed of key recovery and has higher accuracy.
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 412–421, 2019.
https://doi.org/10.1007/978-3-030-15035-8_38
Multiple S-Box Correlation Energy Analysis Model Based on PSO 413
2 Preliminary
2.1 CPA Attack
CPA attack is an extremely effective method of cipher attack [8]. When using this
method to attack a cryptographic device, the attacker does not need to study the device
in detail, and only needs to know the encryption algorithm used. According to the
energy trace generated by the cryptographic device during encryption, the CPA attack
performs key recovery by analyzing the dependence between the energy trace and the
processed data.
From Fig. 1, we can conclude that there are 64 keys in the attack of a single S-box,
so there are 64 curves. In addition to the separated red curve representing the energy
trace produced by the correct key, the remaining curves are the energy traces produced
by the 63 incorrect keys. Through experiments we can conclude that at the point of
difference of about 550 energy curves, we can separate the correct key correlation
coefficient from the error key correlation coefficient.
In Fig. 2, the number of energy traces generated is too large due to the attack of 8
S-boxes as a whole. Therefore, the experimental results show only the correlation
coefficients of some energy curves. The red curve separated from the rest of the curve
represents the energy trace produced by the correct key, and the remaining curves are
the energy traces generated by the partial error key. Through the distinguishing points
in the experiment, we can conclude that for an energy analysis attack of 8 S-boxes, only
250 energy curves are needed to distinguish the correct key from the wrong key.
Compared to the traditional energy analysis attack for a single S-box, the energy
analysis model that attacks 8 S-boxes simultaneously reduces the energy curve required
to distinguish the correct key from the wrong key by about 55%.
416 W. Yao et al.
Step 2: Population initialization. After encoding the key, we needs to initialize the
particle swarm and determine the size and scale of the initial population. In the process
of initializing the population, if the number of birds at the beginning is too small, the
effect of the final solution may be affected, and a local optimal solution may occur; if
the number of birds is too large, the collection speed and solution speed of the bird
group will be reduced. Therefore, the selection of population parameters needs to be
specifically set for specific problems. Table 2 completes the initialization of m indi-
viduals in a population. Each individual in the initial population is equivalent to a
randomly guessed round key.
Step 3: Determine the fitness function and optimize the individual position in the
particle group by calculating the Hamming weight. In each population iteration, you
need to choose a random solution that is closest to the correct key. In this paper, we
chose a discriminant function to calculate the priority of each individual location. The
discriminant function is defined as data input, that is the degree of association between
the Hamming weight and the correct key generated after 8 S-boxes.
The higher the degree of association, the closer the current solution is to the actual
solution, and the better the solution. A new energy analysis model can be established
by calculating the Hamming weight of the eight S-boxes. The individual association
value determines the priority of the individual location. For example, Each generation
has 90 individuals who can solve and rank the associated values for each individual
location. Select the individual location with a high association, and use the particle
velocity and neighborhood definitions to update the historical best position (pbest) and
global best position (gbest) operations.
Step 4: Adjust the moving speed and search range of PSO. In each iteration, the
particle velocity and the locality need to be adjusted to better find a reasonable indi-
vidual. The particle’s velocity update is mainly composed of three parts: its own speed
wvk in the previous iteration; the self-cognitive part c1 r1 ðpbestid xk1
id Þ, indicating the
choice of the particle itself, The distance between the current position of the particle
i and it’s optimal position; the social experience part c2 r2 ðpbestid xk1 id Þ, which
indicates the information sharing and cooperation between the particles, the distance
between the current position of the particle i and the best position of the group.
After each iteration speed and phase adjustment of the group, each particle will
generate a motion vector with the inertia vector under the influence of the regional
optimal solution and the global optimal solution, so that the particle moves to the
optimal position and is correctly aligned. The key is further solved.
Step 5: Align the correlation coefficient [16]. After several iterations, compare the
correlation coefficients. If the preset correlation function value is reached, the optimal
position is obtained, and the correct key is restored. Otherwise return to Step 3 and
optimize the individual again.
During the motion, the PSO will ensure that the guess keys in the 8 S-boxes are all
derived from the correct key. Once the correct S-box key is found, the current solution
would be better than that not reaching the correct S-box key. Therefore, during each
iteration, the individual containing the partial correct S-box key position will step
closer to the optimal position. Due to changes in particle velocity and clinical fields, the
location code represents the key can be optimized for a global optimal solution.
Compared with the traditional energy analysis model, our new energy analysis
model is combined with the artificial intelligence algorithm to perform key recovery.
The model will have a qualitative improvement in efficiency and accuracy. In the
process of recovering the key, the new energy analysis model can make better use of
the energy information and enhance the correlation between the energy consumption of
each S-box, so that the energy curve required is greatly reduced.
418 W. Yao et al.
Fig. 3. Influence of inertia weight and acceleration factor on the number of iterations
From Fig. 3, we can get that under the premise of the initial population fixed, the
inertia weight w and the acceleration factors c1, c2 will affect the number of iterations
required to recover the key [17]. When the inertia weight w = 0.65 and the acceleration
factors c1 and c2 are 1.60, the PSO can achieve better optimization results.
Figure 4 shows the average number of iterations for different population numbers
obtained after 10 calculations when the inertia weight w = 0.65 and the acceleration
factors c1 and c2 are 1.60.
The trend shown in the figure is generated because as the number of populations
increases, the number of population iterations decreases first and then increases,
eventually returning to stability, thereby responding to changes in information and
population diversity in the particle swarm [18].
Figure 6 shows the relationship between the energy curve required to recover the
key and the correct rate in the DES encryption algorithm using the correlation energy
analysis method. From the figure we can see that when the energy curve is 600, the
correct rate of the new energy attack model has reached 100%, while the traditional
energy analysis model has a correct rate of about 70%, and the correct rate is 30%.
Through experiments, it is concluded that under the condition of limited energy curve,
the success rate of recovery key recovery of the new energy attack model combined
with PSO has been improved.
5 Conclusion
We proposes a multiple S-box energy analysis attack model based on particle swarm
optimization in this paper. The model can quickly search the key space by using the
particle swarm optimization algorithm, and convert the key recovery into a binary
string solution, which greatly improves the key recovery speed. In the design and
construction process, the model satisfies the requirements of the efficiency and cor-
rectness of the correlation energy analysis attack theory. Compared with the traditional
correlation energy analysis model, the improved energy analysis model makes the
cryptographic device have a better correlation with the actual energy consumption.
Through experimental analysis and verification, we can conclude that the multi-S box
energy analysis model based on particle swarm optimization reduces the demand for
the number of energy curves by about 55% compared with the traditional single S-box
energy analysis model. In the case of a limited number of curves, the key recovery
correct rate is increased by 30%. It explains that our new energy analysis model can
recover the correct key more efficiently and accurately.
References
1. Brier, E., Clavier, C., Olivier, F.: Correlation power analysis with a leakage model. In:
Cryptographic Hardware and Embedded Systems-CHES 2004 (2004)
2. Li, H., He, G., Guo, Q.: Similarity retrieval method of organic mass spectrometry based on
the Pearson correlation coefficient. Chem. Anal. Meterage 24(3), 33–37 (2015)
3. Mizuno, H., Iwai, K., Tanaka, H., Kurokawa, T.: A correlation power analysis counter-
measure for Enocoro-128 v2 using random switching logic. In: 2012 Third International
Conference on Networking and Computing (ICNC) (2012)
4. Zhang, Z., Wu, L., Wang, A., et al.: Improved Leakage Model Based on Genetic Algorithm.
IACR Cryptology EPrint Archive, 2014: 314 (2014)
5. Nakai, T., Shibatani, M., Shiozaki, M., Kubota, T., Fujino, T.: Side-channel attack resistant
AES cryptographic circuits with ROM reducing address-dependent EM leaks. In: 2014 IEEE
International Symposium on Circuits and Systems (ISCAS) (2014)
6. Qiu, W.-X., Xiao, K.-Z., Ni, F., Huang, H.: DES key extension method. Comput. Eng. 37(5),
167–168+171 (2011)
7. Khalid, S., Khalil, T., Nasreen, S.: A survey of feature selection and feature extraction
techniques in machine learning. In: Science and Information (2014)
Multiple S-Box Correlation Energy Analysis Model Based on PSO 421
8. Zhang, H., Zhou, Y., Feng, D.: Theoretical and practical aspects of multiple samples
correlation power analysis. Secur. Commun. Netw. 9(18), 5166–5177 (2016)
9. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks, pp. 1942–1948. IEEE (1995)
10. Pathak, V.K., Singh, A.K., Singh, R., Chaudhary, H.: A modified algorithm of particle
swarm optimization for form error evaluation. tm-Tech. Mess. 84(4), 272–292 (2017)
11. Huang, W.-X.: Research on the development of particle swarm optimization. Comput. Eng.
Softw. 35(4), 73–77 (2014)
12. Hemanth, D.J., Umamaheswari, S., Popescu, D.E., Naaji, A.: Application of genetic
algorithm and particle swarm optimization techniques for improved image steganography
systems. Open Phys. 14(1), 452–462 (2016)
13. Kennedy, J., Eberhart, R.C.: A discrete binary version of the particle swarm algorithm. In:
1997 IEEE International Conference on Systems, Man, and Cybernetics. Computational
Cybernetics and Simulation (1997)
14. Hu, M.-Y., Liu, R.-H.: Analysis and research on security of DES algorithm. Acta Sci. Nat.
Univ. NeiMongol 6, 95–99 (2005)
15. Pan, Q., Zhang, L., Dai, G., et al.: Two denoising methods by wavelet transform. IEEE
Trans. Signal Process. 47(12), 3401–3406 (1999)
16. Lopez, M.C., Fabregas, X.: Polarimetric SAR speckle noise model. IEEE Trans. Geosci.
Remote Sens. 41(10), 2232–2242 (2003)
17. Wang, D.-F., Meng, L.: Performance analysis and parameter selection of PSO algorithm.
Acta Autom. Sin. 42(10), 1552–1561 (2016)
18. Duan, X.-D., Gao, H.-X., Zhang, X.-D., Liu, X.-D.: Relations between population structure
and population diversity of particle swarm optimization algorithm. Comput. Sci. 34(11),
164–166 (2007)
Improving Peer Reliability Considering
Jitter Parameter: A Fuzzy-Based System
for JXTA-Overlay P2P System
1 Introduction
The Internet is growing every day and the performance of computers is increased
exponentially. However, the Internet architecture is based on Client/Server (C/S)
topology, therefore can not use efficiently the clients features. Also, with appear-
ance of new technologies such as ad-hoc networks, sensor networks, body networks,
home networking, new network devices and applications will appear. Therefore, it
is very important to monitor, control and optimize these network devices via com-
munication channels. However, in large-scale networks such as Internet, it is very
difficult to control the network devices, because of the security problems.
The Peer-to-Peer (P2P) networks will be very important for future dis-
tributed systems and applications. In such systems, the computational burden of
the system can be distributed to peer nodes of the system. Therefore, in decen-
tralized systems users become themselves actors by sharing, contributing and
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 422–432, 2019.
https://doi.org/10.1007/978-3-030-15035-8_39
Improving Peer Reliability Considering Jitter Parameter 423
controlling the resources of the system. This characteristic makes P2P systems
very interesting for the development of decentralized applications [1,2].
In [1], it is proposed a JXTA-based P2P system. JXTA-Overlay is a mid-
dleware built on top of the JXTA specification, which defines a set of protocols
that standardize how different devices may communicate and collaborate among
them. It abstracts a new layer on the top of JXTA through a set of primitive
operations and services that are commonly used in JXTA-based applications and
provides a set of primitives that can be used by other applications, which will be
built on top of the overlay, with complete independence. JXTA-Overlay provides
a set of basic functionalities, primitives, intended to be as complete as possible
to satisfy the needs of most JXTA-based applications.
In P2P systems, each peer has to obtain information of other peers and
propagate the information to other peers through neighboring peers. Thus, it is
important for each peer to have some number of neighbor peers. Moreover, it is
more significant to discuss if each peer has reliable neighbor peers. In reality, each
peer might be faulty or might send obsolete, even incorrect information to the
other peers. If a peer is faulty, other peers which receive incorrect information on
the faulty peer might reach a wrong decision. Therefore, it is critical to discuss
how a peer can trust each of its neighbor peers [3–5].
The reliability of peers is very important for safe communication in P2P
system. The reliability of a peer can be evaluated based on the reputation and
interactions with other peers to provide services. However, in order to decide the
peer reliability are needed many parameters, which make the problem NP-hard.
Fuzzy Logic (FL) is the logic underlying modes of reasoning which are approx-
imate rather then exact. The importance of FL derives from the fact that most
modes of human reasoning and especially common sense reasoning are approxi-
mate in nature. FL uses linguistic variables to describe the control parameters.
By using relatively simple linguistic expressions it is possible to describe and
grasp very complex problems. A very important property of the linguistic vari-
ables is the capability of describing imprecise parameters.
The concept of a fuzzy set deals with the representation of classes whose bound-
aries are not determined. It uses a characteristic function, taking values usually in
the interval [0, 1]. The fuzzy sets are used for representing linguistic labels. This can
be viewed as expressing an uncertainty about the clear-cut meaning of the label.
But important point is that the valuation set is supposed to be common to the
various linguistic labels that are involved in the given problem.
The fuzzy set theory uses the membership function to encode a preference
among the possible interpretations of the corresponding label. A fuzzy set can be
defined by examplification, ranking elements according to their typicality with
respect to the concept underlying the fuzzy set [6].
In this paper, we propose a fuzzy-based peer reliability system for JXTA-
Overlay P2P platform considering four parameters: Data Download Speed
(DDS), Number of Interactions (NI), Sustained Communication Time (SCT)
and Jitter (JTT) to decide the Peer Reliability (PR).
The structure of this paper is as follows. In Sect. 2, we introduce JXTA-
Overlay. In Sect. 3, we present the proposed fuzzy-based peer reliability system.
424 Y. Liu et al.
In Sect. 4, we discuss the simulation results. Finally, conclusions and future work
are given in Sect. 5.
2 JXTA-Overlay
JXTA-Overlay project is an effort to use JXTA technology for building an over-
lay on top of JXTA offering a set of basic primitives (functionalities) that are
most commonly needed in JXTA-based applications [7–11]. The proposed over-
lay comprises the following primitives:
• peer discovery,
• peer’s resources discovery,
• resource allocation,
• task submission and execution,
• file/data sharing, discovery and transmission,
• instant communication,
• peer group functionalities (groups, rooms etc.),
• monitoring of peers, groups and tasks.
This set of basic functionalities is intended to be as complete as possible
to satisfy the needs of JXTA-based applications. The overlay is built on top of
JXTA layer and provides a set of primitives that can be used by other appli-
cations, which on their hand, will be built on top of the overlay, with complete
independence. The JXTA-Overlay project has been developed using the ver-2.3
JXTA libraries. In fact, the project offers several improvements of the origi-
nal JXTA protocols/services in order to increase the reliability of JXTA-based
distributed applications and to support group management and file sharing.
The architecture of the P2P distributed platform we have developed using
JXTA technology has two main peers: Broker and Client. Altogether these two
peers form a new overlay on top of JXTA. The structure of JXTA-Overlay system
is shown in Fig. 1.
Database
Peer Information
Input Parameter
JXTA-Overlay Fuzzy System
Output Parameter
Join Nerwork
Peer
DDS
NI
FPRS PR
SCT
JTT
Table 1. FRB.
4 Simulation Results
In this section, we present the simulation results for our FPRS system. In our
system, we decided the number of term sets by carrying out many simulations.
Fig. 5. Relation of PR with DDS and NI for different SCT when JT = 10.
430 Y. Liu et al.
Fig. 6. Relation of PR with DDS and NI for different SCT when JT = 50.
From Figs. 5, 6 and 7, we show the relation between DDS, NI, SCT, JT and
PR. In these simulations, we consider the SCT and JT as constant parameters.
In Fig. 5, we consider JT value 10 units. We change the SCT value from 10 to 90
units. When the SCT increases, the PR is increased. Also, when the DDS and
NI are high, the PR is high. In Figs. 6 and 7, we increase JT values to 50 and
90 units, respectively. We see that, when the JT increases, the PR is decreased.
We conclude that the JT parameter has a big effect on the peer selection.
Improving Peer Reliability Considering Jitter Parameter 431
Fig. 7. Relation of PR with DDS and NI for different SCT when JT = 90.
References
1. Xhafa, F., Fernandez, R., Daradoumis, T., Barolli, L., Caballé, S.: Improvement
of JXTA protocols for supporting reliable distributed applications in P2P systems.
In: International Conference on Network-Based Information Systems (NBiS-2007),
pp. 345–354 (2007)
2. Barolli, L., Xhafa, F., Durresi, A., De Marco, G.: M3PS: a JXTA-based multi-
platform P2P system and its web application tools. Int. J. Web Inf. Syst. 2(3/4),
187–196 (2007)
3. Aikebaier, A., Enokido, T., Takizawa, M.: Reliable message broadcast schemes
in distributed agreement protocols. In: International Conference on Broadband,
Wireless Computing, Communication and Applications (BWCCA-2010), pp. 242–
249 (2010)
4. Spaho, E., Sakamoto, S., Barolli, L., Xhafa, F., Ikeda, M.: Trustworthiness in P2P:
performance behaviour of two fuzzy-based systems for JXTA-Overlay platform.
Soft Comput. 18(9), 1783–1793 (2014)
5. Watanabe, K., Nakajima, Y., Enokido, T., Takizawa, M.: Ranking factors in peer-
to-peer overlay networks. ACM Trans. Auton. Adapt. Syst. 2(3), 1–26 (2007)
6. Terano, T., Asai, K., Sugeno, M.: Fuzzy Systems Theory and Its Applications.
Academic Press Professional Inc., San Diego (1992)
7. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improvement
of JXTA-Overlay P2P platform: evaluation for medical application and reliability.
Int. J. Distrib. Syst. Technol. (IJDST) 6(2), 45–62 (2015)
8. Liu, Y., Barolli, L., Spaho, E., Ikeda, M., Caballe, S., Xhafa, F.: A fuzzy-based
reliability system for P2P communication considering number of interactions, local
score and security parameters. In: International Conference on Network-Based
Information Systems (NBiS-2014), pp. 484–489 (2014)
9. Spaho, E., Matsuo, K., Barolli, L., Xhafa, F., Arnedo-Moreno, J., Kolici, V.: Appli-
cation of JXTA-Overlay platform for secure robot control. J. Mobile Multimed.
6(3), 227–242 (2010)
10. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L.: Improving reliability of
JXTA-Overlay platform: evaluation for e-learning and trustworthiness. J. Mobile
Multimed. 11(2), 34–50 (2015)
11. Liu, Y., Sakamoto, S., Matsuo, K., Ikeda, M., Barolli, L., Xhafa, F.: Improving
reliability of JXTA-Overlay P2P platform: a comparison study for two fuzzy-based
systems. J. High Speed Netw. 21(1), 27–45 (2015)
ANN Based Intrusion Detection Model
1 Introduction
Network attacks prediction is an important technology of security management.
Existing network intrusion detection and prediction methods did not fully con-
sider the specific environment factors of target network, which may make the
results deviate from the true situation [1]. In addition, the existing intrusion
detection technology uses an intrusion detection technique that detects a traffic
event based on the signature of the attacker and comparing it with the existing
signature. However, as attack techniques evolve and become more sophisticated,
a technique for defending against attacks in real time is becoming necessary. As
diverse attacks occur, the existing pattern-matching based intrusion detection
technology cannot detect attacks with new patterns and has difficulty detect-
ing attacks in real time [2]. Owing to these limitations, the amount of research
on intrusion detection technology for detecting attack patterns based on data
mining is increasing [3].
The performance of intrusion detection systems may be changed through
the selection of a traffic detection model, traffic detection criteria, and detec-
tion factors. In particular, such techniques may use many different resources for
analysis and may result in large numbers of false positives. Therefore, a tech-
nique for reducing the use of resources is required for precise traffic modeling
and analysis that can lower the false-positive rate. In this paper, we propose an
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 433–437, 2019.
https://doi.org/10.1007/978-3-030-15035-8_40
434 S. Park and H. Park
In this paper, we use a recent IDS dataset namely CICIDS2017, which covers all
the eleven necessary criteria with common updated attacks such as Denial-of-
service (DoS), Distributed DoS (DDoS), Brute Force, Cross-site Scripting (XSS),
SQL Injection, Infiltration, Port scan and Botnet. The dataset is completely
labelled and more than 78 network traffic features extracted and calculated
for all benign and intrusive flows by using CICFlowMeter software which is
publicly available in Canadian Institute for Cybersecurity website [4]. Unlike
the existing datasets such as DARPA dataset, KDD Cup 99 dataset [5] and
CAIDA dataset, CICIDS2017 dataset provides 14 different attack methods to
which attack detection techniques can be applied. In particular, it includes var-
ious types of DoS attacks that deplete the server resources, such as DoS Hulk,
GoldenEye, Slowloris, and Slowhttptest, as well as a port scan, which is a typ-
ical network attack, and brute force attacks to capture an authority such as
FTP-Patator and SSH-Patator [6].
This study proposes an intrusion detection and prediction model applying the
machine learning algorithm based on 78 data features using the above-mentioned
CICIDS2017 dataset1 [6,7].
1
https://www.unb.ca/cic/datasets/ids-2017.html.
ANN Based Intrusion Detection Model 435
Parameters Value
Independent variable y
Dependent variables x, {x = 1, ...,39}
Learning rate 1, 0.1, 0.01, 0.001, 0.0001
Optimizer Adam, Adagrad, RMSProp, Gradient Descent, Momentum
Hidden dimension 256, 512
Dropout 1, 0.9, 0.8, 0.7
In this study, the abnormal traffic prediction model applying the deep learn-
ing algorithm is implemented using the CICIDS2017 dataset provided by the
CIC (Canadian Institute for Cybersecurity) as mentioned above, and the model-
ing results are compared [5]. Through deep-learning based modeling, the normal
traffic and attack traffic of the CICIDS2017 data are predicted, and finally, which
attack will occur in the attack traffic is predicted. The experimental environment
used to evaluate the performance of the model is outlined with the GeForce
GTX1070 8 GB for GPU, Python 3.7, and TensorFlow 1.9. The CICIDS2017
dataset is composed of 2,830,743 records, of which 70% is used for training,
20% for evaluation, and 10% for validation. A total of 39 independent variables
are extracted by conducting feature-selection based on the Pearson correlation
coefficient from the 78 features representing the data characteristics. Selected
39 independent variables are below for input dataset: Bwd Packet Length Std,
Bwd Packet Length Max, Bwd Packet Length Mean, Avg Bwd Segment Size,
Packet Length Std, Max Packet Length, Packet Length Variance, Fwd IAT Std,
Packet Length Mean, Average Packet Size, Idle Max, Idle Mean, Flow IAT Max,
Fwd IAT Max, Idle Min, Flow IAT Std, Min Packet Length, Bwd Packet Length
Min, Fwd IAT Total, Flow Duration, FIN Flag Count, PSH flag Count, Flow
IAT Mean, Bwd IAT Std, Fwd IAT Mean, URG Flag count, Destination Port,
ACK Flag Count, Fwd Packet Length Min, Bwd IAT Max, Idle Std, Init Win
bytes backward, Fwd Packet Length Mean, Avg Fwd Segment Size, SYN Flag
Count, Fwd PSH Flags, Down/Up Ratio, Fwd Packet Length Max, and Fwd
Packets/s.
Furthermore, 8 dependent variables for detecting attacks are extracted,
excluding some attacks that have an extremely small dataset only to a traffic
imbalance. Selected 8 dependent variables are below for output dataset: Benign,
DDoS, DoS GoldenEye, DoS Hulk, DoS Slowhttptest, DoS Slowloris, FTP Pata-
tor, and Portscan.
To validate the feature selection and traffic imbalance, the prediction results
are compared with the results using the complete datasets without an extraction
process. The 39 independent variables selected as input are regularized using
StandardScalar, and one-hot encoding is conducted for the 9 dependent variables
selected as output.
436 S. Park and H. Park
4 Experimental Results
To obtain the optimal hyperparameters for ANN modeling, the model is trained
while varying the learning rate and dropout. As a result, the optimal result is
obtained when the learning rate is 0.001 and the dropout is 1. Figure 1 shows
the results of applying the Gradient Descent (GD) and Adam optimizers to the
above hyperparameters. According to the modeling results of an ANN using
the GD optimizer, the prediction accuracy is 98% or higher for most attacks.
However, for FTP-Patator attacks, the sensitivity is 0, which indicates that
the ANN modeling cannot predict certain attacks that show a class imbalance.
Consequently, the results of applying other optimizers are analyzed in addition
to ANN modeling. Although only the optimizer is changed, the class imbalance
problem in certain attacks (e.g., FTP-Patator) disappeared. When the Adam
optimizer is applied to ANN modeling, the prediction ratios are constant for all
attacks in general.
5 Conclusion
In this paper, we proposed an anomaly detection model based IDS using the
ANN learning algorithm. Experimental results on the benchmark CICIDS2017
dataset shows that its performance is comparable to that of the ANN with GD
optimizer and ANN with Adam optimizer. However, since the ANN with Adam
optimizer based IDS model uses only a single hidden layer, its computational
overhead is comparatively less than that of the ANN with GD optimizer based
IDS models. Therefore, the ANN with Adam optimizer based IDS model is an
ANN Based Intrusion Detection Model 437
appropriate candidate for real time deployment and intrusion detection analysis.
For our future work, we aim to fine tune various parameters of the proposed IDS
model to further enhance its performance.
References
1. Roesch, M.: Snort - lightweight intrusion detection for networks. In: Proceedings
13th USENIX Conference on System Administration, pp. 229–238 (1999)
2. Mukkamala, S., Janoski, G., Sung, A.: Intrusion detection using neural networks
and support vector machines. In: Proceedings International Joint Conference on
Neural Networks, pp. 1702–1707 (2002)
3. Yu, Y., Wu, H.: Anomaly intrusion detection based upon data mining techniques
and fuzzy logic. In: Proceedings IEEE International Conference on Systems, Man,
and Cybernetics, pp. 514–517 (2012)
4. CICFlowMeter for network traffic generator and analyser. Canadian institute for
cybersecurity (CIC) (2017). https://www.unb.ca/cic/research/applications.html
5. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.: A detailed analysis of the KDD
CUP 99 data set. In: Proceedings IEEE Symposium on Computational Intelligence
for Security and Defense Applications, pp. 1–6 (2009)
6. Sharafaldin, I., Lashkari, A., Ghorbani, A.: Toward generating a new intrusion detec-
tion dataset and intrusion traffic characterization. In: Proceedings 4th International
Conference on Information Systems Security and Privacy, pp. 108–116 (2018)
7. Shiravi, A., Shiravi, H., Tavallaee, M., Ghorbani, A.: Toward developing a systematic
approach to generate benchmark datasets for intrusion detection. Comput. Secur.
31(3), 357–374 (2012)
The 12th International Workshop on
Telecommunication Networking,
Applications and Systems (TeNAS-2019)
Implementation of a Transnational Testbed
and Web UI System with Layer3 SDX
1 Introduction
Today, many applications and services are developed on the Internet, including those
supported by cloud computing and big data. They are the most popular technologies, so
many people use or study them. To satisfy the increasing number of service requests,
service providers have to extend their infrastructure and build many datacenters across
the world to enhance service capability and redundancy. However, when creating new
datacenters, they encounter problems, such as lack of IP address, management oper-
ations among datacenters, bandwidth assignment, quickly shifting traffic patterns, and
security. To respond to these issues, a novel network architecture may be useful.
Software Defined Network (SDN) is a popular current research topic. Compared
with traditional architecture, SDN is differentiated because the control plane and the
data plane are separated. The architecture of SDN is shown in Fig. 1(a). To implement
the control plane, an SDN controller is installed on a server, for example, ONOS (Open
Network Operating System) [1], OpenDayLight (ODL) [2], and Ryu [3]. The SDN
switches support the data planes. When the first packet enters an SDN switch, it will be
forwarded to the SDN controller. The SDN controller has rules that decide policies and
writes flow entries to SDN switches. Consequently, the subsequent packets can follow
the appropriate policy to arrive at the right destination. Furthermore, the SDN network
provides for network programmability. SDN can rapidly satisfy the requests of network
service deployments and provide dynamic updates about network status, automatic
network provisioning, and network virtualization functions [4–6].
An Internet eXchange Point (IXP) is able to provide a connection between the
Internet Service Provider (ISP) and the Content Delivery Network provider (CDN). It
can ensure connection quality and decrease network connection costs and end-to-end
latency. In [7, 8], the authors apply SDN to IPX architecture, called an SDN eXchange
point (SDX).
In [7], authors propose an SDX architecture which consists of ONOS SDN and an
IPX. In this SDX, an ONOS application, SDN-IP [9], can interconnect IP and SDN
domains seamlessly. While the SDN-IP receives Border Gateway Protocol (BGP) in-
formation from IP domains, it will update the routing information base (RIB) within
ONOS. Then, ONOS is triggered to write transfer rules within SDN switches.
In this paper, we will describe how to deploy SDN-IP in an academic research
network. Furthermore, we will describe an implementation of a Web UI system for
SDN-IP, which can help manage the SDX network.
The remainder of this paper is organized as follows. Section 2 introduces the
research background, including ONOS, SDN-IP, and Quagga [10]. The deployment of
SDX architecture and implementation of the Web UI system for SDN-IP are described
in Sect. 3. Experiments and results are demonstrated in Sect. 4. The conclusions are
given in the last section.
2 Background
In our proposed SDX and system, three applications, such as ONOS, SDN-IP, and
Quagga, are the most important. To explain more clearly, the implementation of the
proposed SDX is introduced here.
Implementation of a Transnational Testbed 443
2.1 ONOS
ONOS is an open-source operating system for an SDN control plane, which supports
the current infrastructure of datacenter virtualization and cloud computing. Moreover, it
is production SDN control plane operating system. ONOS has many applications,
excellent modularization, and support both southbound and northbound APIs. It has
higher scalability, availability, and performance than earlier SDN controllers, such as
NOX [11], POX [12], Beacon [13], and Floodlight [14].
ONOS is distributed as a cluster architecture. If a failure happens to one ONOS
implementation, it will follow pre-defined standard procedures to transfer all data to
other implementations. The Network data in SDN is stored on more than one ONOS
system. Each ONOS implementation has specific tasks to process. Thus, the fault
tolerance mechanism of control plane and network management in the large-scale SDN
network can be enhanced.
The ONOS northbound interface supports the Rest API [15] and Intent [16]
framework. Users can communicate with ONOS by them. For example, if users
implement a REST (Representational State Transfer) API related to devices manage-
ment or an Intent app about routing, they can ask ONOS to manage devices with the
REST API, or add/remove flow entries within SDN switches via the Intent app. The
southbound interface supports P4 [17], OpenFlow, NETCONF [18], TL1 [19], SNMP
[20], and RESTCONF [21] protocols so that ONOS can manage devices that support
those protocols, including, white-box switches, and traditional network devices.
2.2 Quagga
Quagga is an open-source routing software suite under terms of the GNU General
Public License (GPL). It can be installed on Linux as a software-based router. The
architecture of Quagga is shown in Fig. 1(b). It is divided into three parts:
• Routing programs (red blocks): supporting routing protocols, including OSPFv2,
OSPFv4, RIPv1, RIPv2, ISIS, and BGPv4.
• Zebra core program (blue blocks): an abstraction layer between the Linux core and
the router.
• Vtysh (virtual teletype shell) (green blocks): the Telnet server is implemented in
routing programs. When a routing program starts, it will specify a port number to
listen, such as 2605. Therefore, users can login to the server to set the routing
configuration via this port.
The core daemon of Zebra is Zserv, which runs on the server side of Unix or TCP
sockets. It can communicate with a client (denoted as zclient API) on the router to
receive events, such as add, modify, and delete. Zebra runs on the user space of Linux.
By Netlink communication mechanism, it updates routing information recorded within
the routing program to the routing table on kernel space of Linux to ensure all routing
entries in the routing table is up-to-date.
444 W.-Y. Huang et al.
3 Implementation
For this research project, we built an SDN-IP platform. It is a layer3 SDX prototype. In
this section, we illustrate the architecture and how it interconnects with multiple
international research networks. In addition, we describe the implementation details for
the SDN-IP Web UI system.
purposes: SDN-IP information display and SDX network topology. For SDN-IP
information display, it has three types of data as follows:
• BGP-Neighbors: show the BGP information of the local site (SDN-IP) and remote
site (Quagga connected with SDN-IP), such as BGP version, BGP ID, and AS
number.
• BGP-Routes: show all information of the routing table.
• BGP-Speakers: show information of all SDN-IP Speakers, such as speaker name,
VLAN id, device id of attachment point, and peers.
After the Web Server obtains the three types of data, it translates these data into a
JSON-formatted string and transfer to the Web UI system. Next, the Web UI fetches
information from the JSON-formatted string with a JSON Parser coded in JavaScript
and displays it in a tabular format.
For the SDX network topology display, the Web Server fetches the AS path from
BGP-Routes and then fetches data of nodes and links. After the Web Server obtains all
data of nodes and links, it sends these data to the Web UI in a JSON-formatted string so
that the Web UI can display a global topology graph. Figure 5(b) shows the pseu-
docode of the algorithm of fetching these data. We present an example to illustrate this
algorithm. There are 2 SDN-IP sites, and they are A (ASN65000) and B (ASN65111)
respectively. The steps of the algorithm are as follows:
• Step 1. Obtain BGP-Routers data from A site, and obtain the AS path. This AS path
has 2 items and they are [65000] and [65000, 65111] respectively.
• Step 2. We obtain the item [65000] and obtain the ASN65000 from this item.
• Step 3. Since no ASN65000 exists in the nodesArray, it will be added to the
nodesArray.
• Step 4. Obtain the next ASN (denotes as nextASN).
• Step 5. Since nextASN is empty, no link is established.
• Step 6. Repeat step2 to step5. A new node (ASN65111) is added in nodesArray and
a link (ASN65000-ASN65111) is also added in linksArray.
• Step 7. Repeat step 1 to step 6 to fetch nodes and links from the ASN path of B site
until all SDN-IP sites are traversed.
Fig. 5. (a) Execution procedure of web UI system. (b) Pseudocode of algorithm for fetching
nodes and links.
448 W.-Y. Huang et al.
4 Testbed in Web UI
Fig. 6. (a) Portal page. (b) Information page. (c) Topology page.
Implementation of a Transnational Testbed 449
In this paper, we describe an SDX network architecture and visualized platform based
on SDN-IP technology and Web UI system. We deployed an SDX on TWAREN for
international nodes. In the SDX deployment, we deployed ONOS and launched SDN-
IP on it to interconnect IP domains connected with the SDX.
We successfully demonstrated this SDX in a meeting of a research consortium -
Collaborations to Enable Transnational Cyberinfrastructure Applications (CENTRA).
Next, we will use intelligence and automation technology of SDN to build the first
operational level SDX in the country to provide a network architecture for intercon-
nected among heterogeneous domains. In the future, we will deploy Scalable Amplified
Group Environment 2 (SAGE2) [22] on the multinational transnational IP network
domain as the network environment for disaster prevention telemetry and smart city
application. Moreover, we will add functions in the Web UI system for network control
and management so that SDX can be more scalable and flexible.
References
1. ONOS. http://onosproject.org/
2. OpenDaylight. https://www.opendaylight.org/
3. Ryu. http://osrg.github.com/ryu/
4. McKeown, N., Anderson, T., Balakrishnan, H., Parulkar, G., Peterson, L., Rexford, J., et al.:
OpenFlow: enabling innovation in campus networks. ACM SIGCOMM Comput. Commun.
Rev. 38(2), 69–74 (2008)
5. ON Foundation: “Software-defined networking: The new norm for networks,” Open
Networking Foundation, Tech. Rep. (2012)
6. Astuto, B.N., Mendonca, M., Nguyen, X.N., Obraczka, K., Turletti, T.: A survey of
software-defined networking: past, present, and future of programmable networks. IEEE
Commun. Surv. Tutorials 16(3), 1617–1634 (2014)
7. Gupta, A., Vanbever, L., Shahbaz, M., Donovan, S., Schlinker, B., Feamster, N., Rexford, J.,
Shenker, S., Clark, R., KatzBassett, E.: SDX: a software defined internet exchange.
ACM SIGCOMM (2014)
8. Lin, P., Hart, J., Krishnaswamy, U., Murakami, T., Kobayashi, M., Al-Shabibi, A., Wang, K.C.:
Seamless interworking of SDN and IP. ACM SIGCOMM Comput. Commun. Rev. 43(4), 475–
476
9. SDN-IP. https://wiki.onosproject.org/display/ONOS/SDN-IP
10. Quagga Routing Suite. http://www.nongnu.org/quagga/
450 W.-Y. Huang et al.
11. Gude, N., Koponen, T., Pettit, J., Pfaff, B., Casado, M., McKeown, N., Shenker, S.: NOX:
towards an operating system for networks. ACM SIGCOMM Comput. Commun. Rev. 38
(3), 105–110 (2008)
12. McCauley, M.: POX (2009). https://github.com/noxrepo/pox
13. Erickson, D.: The beacon OpenFlow controller. In: Proceedings of the Second ACM
SIGCOMM Workshop on Hot Topics in Software Defined Networking, ser. HotSDN 2013,
pp. 13–18. ACM, New York (2013)
14. Floodlight, an open SDN controller. http://www.projectfloodlight.org/floodlight/
15. REST API. http://www.restapitutorial.com/
16. Intent Framework. https://wiki.onosproject.org/display/ONOS/Intent+Framework
17. P4 Language Consortium. https://p4.org/
18. Network Configuration Protocol (NETCONF). https://tools.ietf.org/html/rfc6241
19. Transaction Language 1 (TL1) wiki. https://en.wikipedia.org/wiki/Transaction_Language_1
20. Simple Network Management Protocol (SNMP). https://www.ietf.org/rfc/rfc1157.txt
21. RESTCONF Protocol. https://tools.ietf.org/html/rfc8040
22. SAGE2. http://sage2.sagecommons.org
The Case Study of Software Build-in Design
Based on Quality Factors and FMEA
1 Introduction
As the software requirements change more and more frequently, the quality standards
and methods for improving software are also evolving as in [4].
The design of considering the issues that might occur in the development/test/
operation stage and incorporating related functions into software architecture in the
early development stage as in [1, 3] is called Built-in Design. The Built-in Design
contains functions such as flexible functional expansion capability, scalable operational
mechanism, and software maintainability, etc. and designer of software should choose
among them with consideration of importance and feasibility to the software and the
schedule of the project.
Based on McCall’s software quality factors and FMEA, we conclude the early design
issues, and propose the Built-in Design which is worth to be designed in the early
stage. We design Built-in Design in three aspects, including Operation Control, Quality
Control, and Software Functionality as shown in Table 2.
We illustrate the Cost of Quality diagram of the two sets of software as shown in
Fig. 3 according to the method of measuring Cost of Quality, CoQ as in [12] and as in
(1), and a and b are coefficients of the project.
This study shows the results that under the same Cost of Failure, there is a sig-
nificant difference of Cost of Quality between software with Built-in Design and
software without it. The software with Built-in Design consumes lower unit Cost of
Prevention and results in about 64% reduction of the sum of the Cost of Quality
relatively.
Due to limited R&D resources, the Built-in Design functions based on FMEA was
designed in advance and implemented in three phases. At completion of each phase the
test data sets are inputted and output results are checked. The cost of checking failure
and removing defect are converted into Cost of Failure to obtain the Cost of Failure
curve as shown in Fig. 3.
The Case Study of Software Build-in Design Based on Quality Factors and FMEA 457
The Cost of Prevention curve as shown on the left chart in Fig. 3 shows that the
Cost of Prevention increases dramatically as software quality being improved. The
quality of software was improved by increasing the coverage of test and audit which is
achieved by increasing the number of test cases or the number of audit items. On the
contrary, the Cost of Prevention curve as shown on the right chart in Fig. 3 shows that
the Cost of Prevention increases much slower than the Cost of Prevention curve on the
left chart as software quality being improved. This is achieved by checking the
rationality and legitimacy of the input data and designing appropriate fault-tolerant
correction or error-handling mechanism accordingly and by incorporating functions of
partial logic segment testing when designing operational functions. Therefore the cost
of verifying multi-data input or complex logic conditions and the cost of auditing
execution results can be reduced.
This study found that through the Built-in Design of auditing and environmental
diagnostic tools, failures can be detected early and eliminated accordingly. Therefore the
operation can be completed within the specified time limit and failures detected in a later
phase and a series of following correction measures can be avoided. By developing a
development procedure and introducing automated continuous-integration mechanism
as in [13], defects of analysis, design, and development can be discovered before being
delivered for testing and deploying, which can reduce the failure rate by 25%.
5 Conclusion
By implementing the software with Built-in Design, we can apply quality assurance
measures as in [14] in the software development stage to improve development effi-
ciency and remove defects efficiently and stabilize software quality. Therefore we can
effectively reduce the cost of quality control measures as in [14] and achieves the goal
of producing high-quality software at lower cost that meets the concept of continuous-
integration.
458 M.-L. Hsieh et al.
This study analyzes Built-in Design required to software and explained it in detail.
Facing the diversified and endless software requirements, it is our goal to maintain the
quality and flexibility of software while satisfying user-requirements fast and
efficiently.
As Information Technology continues evolving, we will apply artificial intelligence
technology in the future to analyze and design the Built-in Design that can automati-
cally adapt the failure mode of environmental resources.
References
1. Markopoulos, P., Martens, J.-B., Malins, J., Coninx, K., Liapis, A.: Collaboration in creative
design: methods and tools, January 2016
2. McDermott, R.E., Mikulak, R.J., Beauregard, M.: The Basics of FMEA, 2nd edn.
Productivity Press, New York (2008)
3. Schultz, D.J.: A comparison of five approaches to software development, January 2000
4. Bourque, P., Fairley, R.E.D.: Guide to the Software Engineering Body of Knowledge. IEEE
Computer Society, New York (2014)
5. Cavano, J.P., McCall, J.A.: A framework for the measurement of software quality.
ACM SIGSOFT Softw. Eng. Notes 7, 133–139 (1978)
6. TM Forum, AI: The time is now, December 2017
7. Project Management Institute, A Guide to the Project Management Body of Knowledge:
PMBOK Guide, 5th edn., Project Management Inst. (2008)
8. Pentti, H., Atte, H.: Failure mode and effects analysis of software-based automation systems,
August 2002
9. Duggal, G., Suri, B.: Understanding regression testing techniques. In: National Conference
on Challenges and Opportunities. Mandi Gobindgarh, Punjab, India, 29 March 2008.
Accessed 3 Jan 2018
10. Kiczales, G., Lamping, J., Mendhekar, A., Maeda, C., Lopes, C., Loingtier, J.-M., Irwin, J.:
Aspect-oriented programming. In: Proceedings of the European Conference on Object-
Oriented Programming (ECOOP), June 1997
11. Laddad, R.: AspectJ in Action, 2nd edn. Manning, Greenwich (2009)
12. Campanella, J.: Principles of quality costs, 3rd edn. American Society for Quality Control,
Milwaukee (1990)
13. Humble, J., Farley, D.: Continuous Delivery: Reliable Software Releases Through Build,
Test, and Deployment Automation, 1st edn. Addison-Wesley, New York (2010)
14. Rose, K.H.: Project Quality Management: Why, What and How, p. 41. J. Ross Publishing,
Fort Lauderdale (2005). ISBN 1-932159-48-7
A Novel Sharing M-Coupons with Lightweight
Computations via Cloud Computing
Abstract. Coupons are useful and widely used in the market place to attract
new customers and keep loyal customers. As more and more wireless devices
are connected to the Internet, many have computed that mobile commerce will
blossom soon, just like e-commerce booming in recent years. In this paper, we
proposed a secure sharing m-coupon scheme that a user can use the mobile
device to share the sharable m-coupons to his/her friend through the cloud. In
the proposed scheme, a customer does not need to download the m-coupon from
cloud. If customer wants to redeem the m-coupon, he just needs to send a
redemption request message to the cloud, and the cloud will help the customer to
transfer the m-coupon to the retiler. So that the customers do not worry about
problem of storing and losing, it is convenient for customers. Our implemen-
tation incorporates only lightweight computations, message authentication code,
and simple hash chains for security.
1 Introduction
Nowadays, electronic commerce is very flourishing over the internet. Hence, many
traditional transactions have been proposed to perform electronically. As more and
more wireless devices are connected to the Internet, many have computed that mobile
commerce will blossom soon, just like e-commerce booming in recent years. In
addition, many marketing activities can be conducted on the mobile platform, such as
Short Message Service (SMS) advertising [1]. The mobile marketing is growing fast
because it is more cost effective, personalized, and real time delivery capabilities.
Therefore, advertisers develop digital coupons on mobile devices that make more
attract customers’ attention than traditional coupons. Juniper Research forecasted that
200 million mobile subscribers globally will use mobile coupons by 2013 [2].
A mobile coupon is defined as “a coupon sent to and stored on a mobile phone that
can be exchanged for a rebate, a monetary discount, etc. at a retailer or merchant when
the customer purchases a product [3].” A mobile user obtains an m-coupon in several
ways, such as downloading from a web site, an SMS message, a Multimedia Message
Service (MMS), or receiving it from a marketer or friend. Experts predicted a higher
usage of mobile coupons, which are stored in the memory of mobile device and
therefore difficult to misplace or forget, compared to their paper-based equivalents [4].
Therefore, the marketer can connect to their target customers more direct through the
mobile devices.
2 Related Work
Traditionally, one coupon is only sent to one user. To increase the redemption rate of
coupons, customer can share the coupons to his friends. Figure 1 shows the protocol of
this paper, which are described in the following. The proposed electronic coupon
scheme consists of six entities as follows:
Manufactures: The manufacturer who produces the goods, sells it to customers, and
authorizes the m-coupon provider to issue m-coupons.
Retailers: The retailer is the owner of the shops selling the promoted products. The
customers buy the promoted products at their online store. The retailer holds an
electronic point-of-sale device which is issued by the manufacture. A tamper-resistant
chip is embedded in to the device and contains the coupon provider’s secret key and the
462 I.-C. Lin and H.-C. Chiang
verification function, such as a smart card or a SIM card. The chip in the device
protects the information stored in a secret area from copy or disclosing [18, 19].
Coupon provider: The coupon provider is the entity that issuing m-coupons to
potential customers through a wireless network. The coupon provider can accept a
registration request from a customer or a friend who has received a sharable m-coupon
from a customer.
Cloud: The cloud is the entity that distributing the m-coupons, which are issued by
the coupon provider, to potential customers or a friend who has received a sharable m-
coupon from a customer through a wireless network.
Customers: The customer is the client that buying the products at the retailer’s on-
line store and having registered in the coupon provider. A customer will receive both
m-coupons and sharable m-coupons from the coupon provider.
Friends: The friend will receive an authentication from the customer. He/She can use
the authentication to register with coupon provider to obtain a sharable m-coupon.
4 Proposed Scheme
4.1 Notations
Before proceeding to our scheme, we first introduce some notations that will be used
throughout the paper in describing the e-coupon system (Table 1).
Table 1. (continued)
Notations Description
p The amount of e-coupons which customer wants to request
q The amount of sharable e-coupons which customer wants to request
Xi =Yj =Nl Position of the Chain X=Chain Y=Chain N
(i ¼ 0 p; j ¼ 0 q; l ¼ 0 q)
H() One-way hash function
∥ The concatenation of two strings
Step 1: The cloud L sends a registration request and his IDL to the coupon provider P.
Step 2: After receiving the message from the cloud L, the coupon provider P uses
his secret key KP and the identity of cloud IDL to compute a secret key of cloud
KL ¼ MACðKP ; IDL Þ. He then sends KL to the cloud L through a secure channel.
(2) The P uses the cloud’s secret key KL and IDC , which received from the C to
compute RC ¼ MACðMACðKL ; IDC Þ.
(3) The P chooses a random number Xp and computes the root value
X0 ¼ Hp Xp .
(4) The P chooses a random number Yq and computes the root value
Y0 ¼ Hq Yq .
(5) The P chooses a random number Nq and computes the root value
N0 ¼ Hq Nq .
(6) The P then use his secret key KP ; PRODUCT; SN; EXD; X0 RL ; RC to compute
SX ¼ MACðKP ; PRODUCT k SN k EXD k X0 k RL k RC Þ.
(7) The P then use his secret key KP ; PRODUCT; SN; EXD; Y0 RL ; N0 to compute
SY ¼ MACðKP ; PRODUCT k SN k EXD k Y0 k RL k N0 Þ.
Step 3: The P sends EKC Nq to C, which is encrypted by C′s secret key.
Step4: The P sends EKL IDC ; ðPRODUCT; SN; EXDÞ; Xp ; SX ; Yq ; SY ; N0 to
L, which is encrypted by L′s secret key.
Step 5: Upon receiving the message from P, the L decrypts the message by using his
secret key KL . And then he stores the value which in the message in his database.
(1) The R uses his secret key SKR to decrypt the message, and then checks whether
the expiration date of the e-coupon is valid.
(2) The R uses Xi and i, which are received from L, to computes the hash value
X0 ¼ Hi ðXi Þ.
(3) The R verifies whether the relation SX ¼ MACðKP ; PRODUCT k SN k EXD
k X0 k RL k RC Þ holds or not, where PRODUCT; SN; EXD; RC ; RL are
received from L, and X0 is computed by R in last step.
Step 4: If the above checks are correct, the R transfers PRODUCT; SN; EXD, Xi , and i
to M to see if it has been used before. Then the M computes the hash value X0 ¼
Hi ðXi Þ and finds the record of the e-coupon according to SN and checks whether the
X0 is correct and Xi has been used before. If the check indicates the m-coupon is valid,
the firm will send an acceptance message to the R and record the e-coupon as used.
Otherwise, if the check indicates the e-coupon is invalid, it means the m-coupon has
been used more than once. Then the M sends a rejection message to the R.
(1) The L computes the hash function l times with Nl to obtain N0 , where
N0 ¼ Hl ðNl Þ.
(2) After that, the L uses his secret key KL , received IDF , and N0 to computes
RF ¼ MACðMACðKL ; IDF k N0 Þ; IDF Þ and checks whether it is equal to RF
which he received. If it is not, L stops the phase.
(3) The L computes the hash function q times with Yq to obtain Y0 , where
Y0 ¼ Hq Yq , and generate Yj .
(4) The L uses his secret key KL and identity IDL to compute RL ¼ MACðKL ; IDL Þ.
(5) Finally, the L encrypts ððPRODUCT; SN; EXDÞ; Yj ; j; SY ; RL ; N0 Þ using R′s
public key PKR to ensure that only the R can obtain the message, then he send it
to the R.
Step 3: After R receiving the redemption request message from L, he acts the
following steps.
(1) The R uses his secret key SKR to decrypt the message, and then checks whether
the expiration date of the e-coupon is valid.
(2) The R uses Yj and j, which are received from L, to computes the hash value
Y0 ¼ Hj Yj .
(3) The R verifies whether the relation SY ¼ MACðKP ; PRODUCT k SN k EXD k
Y0 k RL k N0 Þ holds or not, where PRODUCT; SN; EXD; RC ; RL are received
from L, and Y0 is computed by R in last step.
Step 4: If the above checks are correct, the R transfers PRODUCT; SN; EXD, Yj , and j
to M to see if it has been used before. Then the M computes the hash value Y0 ¼
Hj Yj and finds the record of the e-coupon according to SN and checks whether the
Y0 is correct and Yj has been used before. If the check indicates the e-coupon is valid,
the firm will send an acceptance message to the R and record the e-coupon as used.
Otherwise, if the check indicates the m-coupon is invalid, it means the e-coupon has
been used more than once. Then the M sends a rejection message to the R.
5 Security Analysis
5.1 Authentication
Our proposed scheme achieves this requirement by allowing a retailer to make sure the
validity of e-coupons. In the e-coupon issuing phase of our scheme, a valid e-coupon
only generates by the genuine coupon provider with the secret key. Any other user in
this system cannot generate a valid m-coupon without the knowledge of the secret key
of the genuine coupon provider. In addition, if a customer attempt to forge an
m-coupon on behalf a coupon provider is impossible, because the forger difficult to get
valid SX or SY .
In the issuing phase, the coupon provider signs PRODUCT, SN, EXD, X0 (or Y0 ),
RL , and RC (or N0 ) by using the secret key of coupon provider to generate SX or SY ,
which is used for verification during redemption. If a malicious user attempts to redeem
A Novel Sharing M-Coupons with Lightweight Computations via Cloud Computing 467
a forged m-coupon, the verification will fail. Because the forged m-coupon would have
been generating by using illegal versions Xp and Yq . In the redemption phase, when the
retailer received illegal Xi the retailer will compute the one-way hash function i times
and compare the output with the root value X0 in the SX . Therefore, a retailer can detect
the validity of m-coupon by SX (or SY Þ in this system. In addition, in the redemption
phase for sharable m-coupons, if an attacker attempts to intercept the message to obtain
the Nl and l, he cannot succeed. Due to the Nl and l just can be used for only once and
the cloud will check it with the identity of friend, so the entity of friend does not worry
about they may be intercepted or copy by attackers.
Furthermore, a retailer also can verify whether the m-coupon received from an
assigned customer. In the redemption phase, the RC (or RF ) is generated from customer
by using his/her secret key KC (or KF ) and identity IDC (or IDF ), and transfer it to the
cloud. When cloud received the RC (or RF ), he can verify it by using his secret key KL
and IDC . We can assume the secret key KC (or KF ) and KL are issued by a trust party.
Since the secret key KC ¼ MACðKL ; IDC Þ (or KF ¼ MACðKL ; IDF ) is generated with
the secret key of KL , so that the corresponding cloud can verify the validity of identity
of customer. In addition, when the retailer received the redemption message from the
cloud, he also can verify the identity of cloud and customer through SX (or SY Þ with the
KP , which is stored in the point-of-sale device. Therefore, our proposed scheme can
detect any forged or false m-coupon successfully.
5.2 Integrity
In the electronic environment, message can be modified easily especially in the
transaction processes. Therefore, when a retailer receives an m-coupon and its asso-
ciated S_X (or S_Y), we must to ensure the received m-coupon has not been modified.
We assume that a malicious customer or cloud attempts to modify PRODUCT to
change the product information or modify EXD to extend the expiration date. The
modification will not pass the verification of S_X (or S_Y) in the redemption phase.
Therefore, our proposed scheme also can verify the integrity of received m-coupons.
hash function can be used to decide the computation cost. According to the ISO
standard of a MAC algorithm [16], the MAC algorithm encrypts the hash value, which
is computed using the MD5 hash function, by using DES encryption function.
Therefore, MAC’s computation cost is one hash function and one DES encryption
function computation. Table 2 shows the number of public key signature, symmetric
key encryption, and one way hash function that can be performed on a typical work-
station [20, 21]. According to the results, we find that the one-way hash function and
the symmetric key encryption are much cheaper than the public key signature.
Table 3 lists all of the operations in each phase except for the registration phase.
We can see that the proposed schemes indeed require very low computational cost.
Furthermore, the computational capability of the customer’s mobile phone is very
small. Therefore, our proposed scheme can reduce the load of customer’s mobile
phone, and it is practical in the mobile commerce.
Table 3. Analysis of the computational cost for the entities in the proposed scheme.
Phase Entity Operation Times
Registration Customer MAC 0
One-way hash 0
Encryption/Decryption 0
Issuing Customer MAC 0
One-way hash 0
Decryption 1
Redemption Customer MAC 1
One-way hash 0
Encryption 1
Sharing Customer MAC 0
One-way hash 0
Encryption/Decryption 0
Friend MAC 0
One-way hash 0
Encryption/Decryption 0
Redemption for Friend MAC 1
sharable m-coupons One-way hash 0
Encryption/Decryption 0
A Novel Sharing M-Coupons with Lightweight Computations via Cloud Computing 469
6 Conclusions
In this paper, we have presented a secure sharing m-coupon scheme that a user can use
the mobile device to share the sharable m-coupons to his/her friend through the cloud.
Our proposed scheme combines the popularity of cloud to improve coupon redemption
rate, it may benefit both customer and markets. In addition to achieving the three
security requirements- (1) authentication, (2) integrity, and (3) single use, the proposed
scheme also provides (1) low computational cost, and (2) low communication cost. It
can not only be used on the Internet, but also suitable for being used on wireless
networks. Therefore, the proposed scheme is very efficient and practical.
References
1. Scharl, A., Dickinger, A., Murphy, J.: Diffusion and success factors of mobile marketing.
Electron. Commer. Res. Appl. 4(2), 159–173 (2005)
2. Neil, D.S.: Coupons: Identifying New Opportunities Beyond Early Trials. Jupiter Research
Report (2008)
3. Juniper Research: Mobile Coupons (2008). http://www.juniperresearch.com
4. Raskino, M.: Mobile coupons will reach right into your pocket. GartnerGroup Research Note
(2001). http://www.gartner.com/
5. Kumar, M., Rangachari, A., Jhingran, A., Mohan, R.: Sales promotions on the internet. In:
Proceedings of the 3rd USENIX Workshop on Electronic Commerce, pp. 167–176 (1998)
6. Blundo, C., Cimato, S., De Bonis, A.: A lightweight protocol for the generation and
distribution of secure e-coupons. In: Proceedings of the 11th International Conference on
World Wide Web, Honolulu, Hawaii, USA, pp. 542–552 (2002)
7. Blundo, C., Cimato, S., De Bonis, A.: Secure E-coupons. E-Commer. Res. 5(1), 117–139
(2005)
8. Chang, C.C., Wu, C.C., Lin, I.C.: A secure E-coupon system for mobile users. Int.
J. Comput. Sci. Netw. Secur. 6(1), 273–279 (2006)
9. Wray, J., Plante, D.: Mobile advertising engine for centralized mobile coupon delivery. Int.
J. Manag. Mark. Res. 4(1), 75–85 (2011)
10. Hsueh, S.C., Chen, J.M.: Sharing secure m-coupons for peer-generated targeting via eWOM
communications. Electron. Commer. Res. Appl. 9(4), 283–293 (2010)
11. Ricest, R.L., Shamir, A.: PayWord and MicroMint: two simple micropayment schemes. In:
Proceedings of 1996 International Workshop on Security Protocols, pp. 69–87 (1996)
12. Damgard, I.B.: A design principle for hash functions. In: Advances in Cryptology-CRYPTO
1989. LNCS, vol. 435, pp. 416–427 (1990)
13. Merkle, R.C.: One way hash function and DES. In: Advances in Cryptology-CRYPTO
1989. LNCS, vol. 435, pp. 428–446 (1989)
14. Merkle, R.C.: A fast software one-way hash function. J. Cryptol. 3(1), 43–58 (1990)
15. Davies, D.W.: A message authentication algorithm suitable for a main-frame computer. In:
Advance in Cryptology-CRYPTO 1984. LNCS, vol. 196, pp. 393–400 (1984)
16. ISO/IEC 9797: Data cryptographic techniques-data integrity mechanism using a crypto-
graphic check function employing a block cipher algorithm. Internal Organization for
Standardization
17. Menezes, A., Van Oorschot, P., Vanstone, S.: Handbook of Applied Cryptography. CRC
Press Inc., Boca Raton (1996)
470 I.-C. Lin and H.-C. Chiang
18. Hendry, M.: Smart Card Security and Applications. Artech House, Norwood (1997)
19. Jones, H.W.E., Watson, A.C., O’Neill, T.J.: Vehicle security using smartcards. Secur. J.
13(2), 79–87 (1998)
20. Schneier, B.: Applied Cryptography, 2nd edn. Wiley, Hoboken (1996)
21. Baldonado, M., Chang, C.-C.K., Gravano, L., Paepcke, A.: The stanford digital library
metadata architecture. Int. J. Digit. Libr. 1, 108–121 (1997)
A Smart Roadside Parking System Using
Bluetooth Low Energy Beacons
Abstract. The management of roadside parking has been a big issue for
metropolitan areas. Many city governments still rely on manual issuing of
parking tickets or use parking meters to collect the parking fee. Either way, it is
not economical in terms of the expenses on human labor and the efficiency of
parking space management. To solve the problem, some researchers have pro-
posed roadside parking systems based on wireless sensor networks. However,
the existing approaches focus only on detecting the occupancy of parking
spaces. This is not enough for a comprehensive roadside parking system because
the identification of vehicles using specific roadside parking spaces is yet
unknown. In this work-in-progress research, we propose a roadside parking
system based on the Bluetooth Low Energy (BLE) beacon technology to
achieve the two goals simultaneously. Specifically, vehicles are attached with a
BLE beacon device which broadcasts the identification of the registered vehicle
to be scanned by the roadside units. By monitoring the estimated distances
between the beacon device and a number of roadside units, our backend system
is expected to determine the occupancy of the roadside parking spaces along
with the identification of vehicles using the parking spaces.
1 Introduction
the parking spaces regularly, and then issue parking tickets accordingly. Obviously,
this traditional approach may consume a large amount of human labor and time. To
reduce the manpower needs, many city governments use parking meters to collect
parking fees. Although this approach automates the process of fee collection, it only
solves part of the problem in roadside parking management – the management
authority is still unable to know the real-time status (occupied or not) of each parking
space. For a driver looking for an unoccupied parking space, the driver needs drive
along the street to identify an available one. This is inefficient in terms of time spent
looking for a parking space, and introduces unnecessary carbon emissions.
With the latest ICT technologies, many researchers proposed novel approaches to
build a smart roadside parking management system. For instance, in 2013 Zhang et al.
[2] proposed a street parking system based on wireless sensor networks. They deployed
magnetic sensor nodes on every parking space and designed a vehicle detection
algorithm to detect a parking car. In 2014 Barone et al. [3] proposed the intelligent
parking assistant (IPA) architecture where RFID readers and inductive magnetic loops
are deployed on every parking space to detect whether the space is available or
occupied. They even allowed drivers to reserve a parking space at a specific destination
area before the driver departs for the destination. However, we found that with these
approaches, the management system can merely monitor the status of the roadside
parking spaces. More specifically, the management system has no idea of the identi-
fication of the vehicles using the parking spaces, which makes it difficult to bill the
vehicle owners automatically. To solve this problem, some vendors have offered their
solutions such as the smart parking meter developed by AcerITS [4], a Taiwan-based
company. The smart parking meter has a built-in camera which can recognize the
license plate and then uses it for billing. However, license plate recognition cannot
achieve a perfect accuracy of vehicle identification under poor illumination conditions
or obstruction of license plates.
Based on the above observations, in this research we would like to take advantage
of a popular and low-cost technology – Bluetooth Low Energy (BLE) beacons, to build
a smart roadside parking management system. As long as each vehicle is attached with
a registered beacon device, we expect that our system can detect the occupancy of
every roadside parking space and accurately acquire the identification of the vehicle
occupying a specific parking space. Note that this is a work-in-progress research, in
which we are still working on the problem of distance estimation based on radio
signals.
The rest of this paper is organized as follows. In Sect. 2 we introduce the back-
ground knowledge for the Bluetooth beacon technology. In Sect. 3 the system archi-
tecture along with the principle of operations are described. Section 4 shows the
prototype implementation, which includes how we acquire the identification of vehicles
and how to detect the occupancy of a roadside parking space, as well as how we deploy
the Bluetooth network to support the application-layer protocol – MQTT. Finally, the
conclusions and future work are given in Sect. 5.
A Smart Roadside Parking System Using BLE Beacons 473
2 Background Knowledge
In this section we will introduce the Bluetooth beacon technology. According to the
Bluetooth Core Specification version 4.1, the maximum size of a BLE advertising
packet data unit (PDU) is 39 bytes, in which the advertising data is up to 31 bytes. The
advertising data field is exactly where various beacon protocols used to carry the
beacon data. Although the latest Bluetooth specification (version 5.0) has extended the
maximum advertising channel PDU size to 257 bytes, in Bluetooth 5.0 the so called
“legacy advertising PDUs” are still up to 39-byte long in order to make them backward
compatible.
There are a number of Bluetooth beacon pseudo-standards in the market, among
which the most popular ones are Apple’s iBeacon [5], Google’s Eddystone [6], and
Radius Networks’ AltBeacon [7]. As shown in Fig. 1(a), the iBeacon packet format
consists of four fields: 16-byte UUID, 2-byte Major, 2-byte Minor, and 1-byte Tx
Power, where UUID stands for Universally Unique Identifier. According to [5], a
UUID should be specific to an application/deployment use case, and Major/Minor
values are used to hierarchically specify sub-regions within a larger region or use case.
As for the meaning of Tx Power, it is defined as the wireless signal strength at a
distance of 1 m away from the beacon transmitter. By comparing the received signal
strength indicator (RSSI) and the value of Tx power embedded in the beacon packet, a
beacon receiver can estimate its proximity to the beacon transmitter.
Fig. 1 Packet formats of three popular Bluetooth beacons: (a) iBeacon, (b) Eddystone-UID, and
(c) AltBeacon.
The packet formats of Eddystone-UID and AltBeacon are shown in Fig. 1(b) and (c),
respectively. Note that Eddystone actually defines four packet types: Eddystone-UID,
Eddystone-EID, Eddystone-URL, and Eddystone-TLM. Since we focus on identifying
vehicles by UUIDs, we only introduce Eddystone-UID here. In Eddystone-UID we see
474 H.-T. Chen et al.
the 10-byte Namespace and the 6-byte Instance fields, which can be combined to form a
16-byte universally unique beacon ID. Google recommended two methods to generate a
unique 10-byte namespace: a truncated hash of owner’s FQDN, or an elided UUID [8].
The field of Ranging Data is designed to carry the calibrated Tx power at 0 m, which can
be obtained by measuring the actual signal strength from 1 m away and then add
41 dBm to the value. As for AltBeacon, we can see the 20-byte Beacon ID and the 1-
byte Ref RSSI fields. For interoperability purposes, the AltBeacon specification indicates
that the first 16 bytes of the Beacon ID should be unique to the advertiser’s organiza-
tional unit. The Ref RSSI field is used to represent the average received signal strength at
1 m from the advertiser, which is of the same definition as the Tx power field in iBeacon.
To sum up, the above three types of beacon packets serve a common purpose: carrying a
beacon ID as an identification and a transmit power value for proximity estimation.
3 System Design
3.1 System Architecture
The proposed system architecture is depicted in Fig. 2. There are mainly four roles in
our system: beacon transmitter, beacon reader, gateway, and backend server. We
assume that each vehicle has a registered beacon transmitter attached to the right side
mirror (for right-hand traffic countries). The beacon transmitter emits beacon packets
regularly, which then can be detected by the beacon readers deployed at every corner of
the parking spaces along the curb. A gateway device playing the Bluetooth central role
maintains a connection to each beacon reader playing the Bluetooth peripheral role. In
order to run application-layer protocols in the Bluetooth network, we need the gateway
device and the beacon readers capable of running the IP protocol. As a result, we
deployed the 6LoWPAN technology in the Bluetooth network, where the gateway
serves as the 6LoWPAN Border Router (called 6LBR), and all the readers serve as the
6LoWPAN Nodes (called 6LN). Each reader uses the received beacon packets to
identify the vehicles and to estimate the distances between the reader and the nearby
vehicles. Finally, the readers send the vehicle IDs and the estimated distances to the
backend server for parking status analysis.
Fig. 2 System architecture for roadside parking management using BLE beacons.
associated with different reader IDs. To determine the parking space occupied by a
specific vehicle, the server can simply sort the tuples of the same vehicle ID by the
estimated distances. We may use the scenario in Fig. 2 to give an example, where the
vehicle is on the parking space with ID 02. Let’s assume the vehicle ID of this vehicle
is UUID1. All the four readers detected UUID1 and sent their estimated distances to the
backend server. After the server sorted all the tuples by the estimated distances (the
example shown in Table 1), the server found that the vehicle is closest to Reader2,
followed by Reader3. Meanwhile, the server further found that the top-2 distances
(namely, d1 and d2) are below some predetermined thresholds (e.g., d1 < 4 and d2 < 6).
Finally, the server can conclude that the vehicle is on the parking space between
Reader2 and Reader3 (i.e., the parking space 02).
4 Prototype Implementation
In our prototype system, the gateway and the readers are implemented by using
Raspberry Pi 3 development boards which have built-in Bluetooth v4.1 interfaces, and
the beacon transmitter is implemented by using an Estimote Proximity Beacon device
which supports both iBeacon and Eddystone BLE beacon packets. Currently, we
configure the beacon device to emit iBeacon packets in order to test the functionality.
In the following we will describe the process of distance estimation carried out by
beacon readers, and the way we deploy the 6LoWPAN-based Bluetooth network.
To estimate the distance between a beacon reader and a beacon transmitter, we need
the following two pieces of information: the RSSI of the received beacon packet and the
Tx power designated in the beacon packet. In fact, both pieces of information are
available from the previous execution output. From Fig. 4 we can see the value
A Smart Roadside Parking System Using BLE Beacons 477
of −67 dBm, which is the RSSI of the detected iBeacon packet. The value of Tx power
in 2’s complement is 0xC5 (i.e., −59 dBm in decimal), which is located at the last byte
of the Manufacturer field. Then we implemented two different formulas to calculate the
estimated distance d. The first formula is from [10] as shown in (1):
ARSSI
d ¼ 10 10 n ð1Þ
Then, we modified the source code of blescan again by adding the two formulas
to do distance estimation. Figure 5 shows one of the experimental results, where
Distance_M1 and Distance_M2 refer to the distance estimation by using (1) and (2),
respectively. In fact, we have evaluated the accuracy of the distance estimation at 8
different distances ranging from 1 m to 8 m (incremented by 1 m), and at each distance
we collected the distance estimation for 20 times. However, we found that the errors of
both formulas are quite large due to the significant fluctuation of radio signals. We are
still working on this issue and would design a numerical algorithm to smooth the
collected data. As it has been pointed out in the official blog of the Bluetooth SIG [12],
the author suggested that one should avoid using the absolute value of the RSSI – use
the trend instead. Therefore, we will try to take advantage of the trend of the RSSI to
validate the applicability of our proposed approach.
478 H.-T. Chen et al.
Fig. 5 Output of distance estimation between the beacon transmitter and the reader.
Fig. 6 IPv6 configuration of the Bluetooth interface and the ping test at a 6LN.
A Smart Roadside Parking System Using BLE Beacons 479
Next, we need to design the application architecture that can send the distance
estimation data back to the backend server in a timely manner. We choose the popular
application-layer protocol suitable for IoT environments – MQTT [14], to implement
the Publish/Subscribe messaging model. As shown in Fig. 7, beacon readers play the
role of publishers which publish a specific topic containing the distance estimation data
to the gateway acting as a broker. The backend server is configured to subscribe to this
specific topic. Whenever a new piece of distance estimation data is published by any
reader, it will be pushed to the backend server automatically. Consequently, the
backend server can successfully collect the needed data and then perform parking status
analysis. The whole MQTT-based messaging system is implemented using the open-
source Really Small Message Broker (RSMB) [15].
drivers to confirm their parking statuses. Last but not least, to prevent the problem of
counterfeit vehicle IDs, the beacon ID should be protected by some security
mechanisms.
References
1. Smart Taipei - Smart street light empirical field. https://smartcity.taipei/project/120
2. Zhang, Z., Li, X., Yuan, H., Yu, F.: A street parking system using wireless sensor networks.
Int. J. Distrib. Sens. Netw. 9, 1–10 (2013)
3. Barone, R.E., Giuffre, T., Siniscalchi, S.M., Morgano, M.A., Tesoriere, G.: Architecture for
parking management in smart cities. IET Intell. Transport Syst. 8, 445–452 (2014)
4. Acer ITS - On Street Parking - Smart Parking Meter. http://www.acerits.com/index-en.
html#works
5. Apple iBeacon. https://developer.apple.com/ibeacon/
6. Eddystone. https://developers.google.com/beacons/eddystone
7. AltBeacon. https://altbeacon.org/
8. Eddystone-UID. https://github.com/google/eddystone/tree/master/eddystone-uid
9. bluepy - Python interface to Bluetooth LE on Linux. https://github.com/IanHarvey/bluepy
10. Wang, Y., Xu, Y., Yutian, Z., Yue, L., Cuthbert, L.: Bluetooth positioning using RSSI and
triangulation methods. In: 2013 IEEE 10th Consumer Communications and Networking
Conference (CCNC), pp. 837–842 (2013)
11. AltBeacon Android Beacon Library. https://github.com/AltBeacon/android-beacon-library
12. Proximity and RSSI. https://blog.bluetooth.com/proximity-and-rssi
13. BlueZ - Official Linux Bluetooth protocol stack. http://www.bluez.org/
14. MQTT Version 3.1.1. http://docs.oasis-open.org/mqtt/mqtt/v3.1.1/os/mqtt-v3.1.1-os.html
15. RSMB: Really Small Message Broker. https://github.com/eclipse/mosquitto.rsmb
A Light Weight Stream Cypher Mechanism
for Visible Light Communication
Abstract. LED light has many advantages such as power efficient, better
quality, long lifespan, and focus emission. LED lighting has been employed in
our daily life in various ways, such as car lamps and lights in vehicles, roadside
signal lights, indoor lighting systems, etc. The LED light bulbs not only can
provide bright but also can be employed for data communication and object
positioning, namely visible light communication (VLC). VLC is fast-growing
technologies that enable new visible light application such as fast data trans-
mission and indoor positioning. Due to its attractive functionalities, they have
draw lots of attention to the wireless communication and indoor navigation
systems. Because of its physical characteristics, “what you see is what you
send”, VLC has been always considered as a secured communication method. In
this work, we analyze the risks of visible light communication signal such as
signal jamming, data snooping and modification. In order to effectively protect
VLC data communication problems, we design a lightweight stream cypher data
communication through the VLC signal to achieve the secure of the data
communication. Experimental results of the proposed mechanism show that the
WG-8 encryption algorithm can be executed on Raspberry Pi3 Model B, which
has limited computing power. The evaluation result shows WG-8 support
embedded system and CPU utilization reach up to 49%.
1 Introduction
Using LED lighting has many advantages such as power efficient, better quality, long
lifespan, and focus emission. LED lighting has been employed in our daily life in
various ways, (car lamps and lights in vehicles, indoor lighting systems, etc.). Visible
Light Communication (VLC) is a fast-growing technology to provide data communi-
cation using low-cost and energy efficient light-emitting diode (LED) and photodiodes.
Due to its attractive functionalities, it has draws lots of attention to the wireless
communication and navigation systems that also reduces cost after system employed.
This white visible light source can be modulated at high-speed so there is the potential
to transmit data as well as provide illumination, e.g., the visible light communications
consortium [1], the IEEE task group, 802.15.7 [2], standardizing VLC for personal area
network etc. Since history likes to repeat itself a common mistake in the development
of novel communications techniques was to neglect or downplay the security issues:
such was the case with the internet protocol suite, fiber-optics networks and etc.
One of the features in which VLC techniques are considered superior to traditional
wireless communication is security. High directivity and low impermeability of visible
light signal are considered to provide a secure way to transmit data between two peers,
making the data difficult to intercept from outside. The common slogan summarizing
VLC security features is: “What you see is what you send” – WYSIWYS [3]. In this
paper we address the classic security triad of: confidentiality, authenticity and integrity
in VLC. Although block cryptography is more widely used than streaming cryptog-
raphy, however, most of the visible light communication required in terms of time
urgent. Stream encryption is more instantaneous than block encryption, and consumes
less power. Because the data encryption by the block cypher will cause more resources
and increasing complexity. It is not suitable for the block cypher system in visible light
communication, and the stream encryption method for bit-based encryption is much
suitable for use in visible light communication.
To address this problem, we proposed a reliable VLC based data communication
mechanisms. First, link quality-monitoring agents, which will able to identify the
faulty, link which enabling receiver node to notify its transmitter node and resume
connectivity after a link breakage. Second, a reliable visible light transmission system
has been proposed, includes three major modules including environment monitoring,
intelligent control and high speed transmission. The rest of the chapter is organized as
follows. In Sect. 2, background and related works are briefly described. In Sect. 3, we
present our Link Guard Scheme (LGS) and its detail mechanisms. Finally, our con-
clusions and future work are explored in Sect. 5.
transmitted at high speeds using LED light. However, since the link can cause link
broke due to movement such as object slight movement or rotation of the receiver unit.
The poor pointing of the VLC devices may cause the decrease of signal quality or even
link disconnection. Moreover, VLC can be highly directional, it is difficult to re-
establish link that has been lost due to movement or rotation of one of the devices in the
link [4–6]. As such, it is necessary and important to have link recovery as well as
handover mechanisms for VLC systems, yet the fast link recovery is very challenging.
Encryption [7] is a mathematical algorithm used in cyber security for data pro-
tection. Usually there are keys and texts and cipher texts. The interception is just a
cipher text that cannot be understood. When the receiver receives the cipher text, it is
transferred back to the text by the decryption algorithm. If the encryption and
decryption keys are the same, then called symmetric encryption algorithm [8, 9]. If
keys are different, is called an asymmetric encryption algorithm [10, 11]. In the current
encryption algorithm, which is mainly used for cyber security. It can be divided into
two parts, block encryption [12, 13] and stream encryption [14, 15]. Block cypher
encryption is an encrypt method using the same block cipher key and encrypt one piece
a time to guarantee data security. The block cipher itself can only encrypt a single piece
of data whose length is equal to the length of the cipher block.
The studies of encryption mode [16, 17] have included the integrity protection of
data, that is, the error propagation characteristics of the password in the case where
some data is modified. Subsequent researches [18] identified integrity protection as
another completely different cryptographic goal that was not related to encryption. Part
of the modern working model combines encryption and authentication in an efficient
way, called the authentication encryption mode. Although the working mode is usually
applied to symmetric encryption, it can also be applied to public key encryption.
Stream ciphers (or their principles) have been presented in the world of securing
information for many years. Russian encryption machine called Fialka [19] is an
example of a mechanical ciphering machine used in Cold War. Its design resembles a
stream cipher, because for each letter of plaintext a different encryption, dependent on
previous ones, is used (in other words, two same letters at a different position in a
plaintext are likely to encrypt to two different letters in ciphertext).
Dragon [20] is a new word based stream cipher submitted to the Euro-
pean ECRYPT Stream Cipher Project. It is selected as a Focus Phase 3 candidate as of
April, year 2007. The Dragon is designed with both security and efficiency in mind to
satisfy the need for lightweight algorithms, dedicated to hardware environments where
the available resources are heavily restricted. This stream cipher is constructed on a
1024 bits word-based NLFSR, a state update function (F function) and a 64 bit memory
(M). It is able to generate 64-bit of pseudorandom keystream per round, operating on
key sizes of 128 and 256 bits. The F function plays its role in both key setup and 32-bit
output keystream generation. It is diffused by using a network of modular and binary
addition through G and H function. This design constructed large internal states of
1024-bit and expected period of 2576.
484 S.-H. Chang et al.
Streaming encryption is a bit level encryption algorithm. Both the encryption and
the decryption use the same pseudo-random stream as the key. Pseudo-random code
[21] means the randomness of the number. However, because the number of cycles is
too large, the size of the sequence of a pseudo-random number column in one cycle is
greater than 1040, so the loop value cannot be seen [22]. Each time the plaintext data is
encrypted correspondingly with the key data stream, a cipher text data stream is
obtained. To encrypt variable length data, the data must first be divided into separate
cipher blocks. In general, the last piece of data also needs to use the appropriate
padding method to extend the data expansion to the length of the cipher block size.
Fialka M-125 [23] is an Enigma based rotor-cipher machine used during the Cold
War. It is considered one of the most secure cipher machines. Advantages of this cipher
are based on the elimination of the Enigma’s known weaknesses. There are no known
attacks on this cipher. The author carried out performance and statistical tests of three
versions of their construction: 4-bit version, 5-bit version and 8-bit version. They
investigated the number of rotors needed to successfully pass the statistical NIST-
testing suite. Their results show that the 4-bit version with 10 rotors and modular
addition/subtraction passes the tests.
In this paper, we present a stream cipher WG-8 [24], which is a lightweight variant
of the well-known WG stream cipher family as submitted to the eSTREAM project.
WG-8 inherits good randomness properties of the WG stream cipher family such as
period, balance, ideal two-level autocorrelation, ideal tuple distribution, and exact
linear complexity. Moreover, WG-8 is able to resist the most common attacks against
stream ciphers including algebraic attack, correlation attack, differential attack, cube
attack, distinguish attack, discrete fourier transform attack, and time-memory-data
tradeoff attack, thereby providing adequate security for lightweight embedded appli-
cations. Lightweight cryptographic primitives are essential for securing pervasive
embedded devices like RFID tags, smart cards, and wireless sensor nodes. WG-8
stream cipher family with 80-bit secret key and 80-bit initial vector (IV), which can be
regarded as a nonlinear filter generator over finite field F28 . More detail will be
described in the next section.
where Hd(0) and Href(0) are the DC channel gain of the direct and reflected paths,
respectively and Pt is total optical power transmitted by LEDs. The DC channel gain of
the direct path is given as:
8 ðm þ 1Þ
< 2pd2 Adet cosm ðhr ÞTs ð#Þgð#Þ cosð#Þ;
Hd ð0Þ ¼ for 0 # #FOV ð2Þ
:
0; for # [ #FOV
where m is the order of the Lambertian radiant, d is the distance between transmitter
and receiver, hr is the angle of irradiance, 0 is the angle of incidence, Ts (0) is the
optical filter gain, g(0) is the optical concentrator gain and #FOV is the FOV of the
receiver. The DC gain of the reflected path can be written as:
8 ðm þ 1Þ
> 2 qAdet dAw cos ðhr Þ cosðaÞ
m
>
< 2ðpd1 d2 Þ
Hd ð0Þ ¼ cosðbÞTs ð#Þgð#Þ cosð#Þ ð3Þ
>
: for 0 # #FOV
>
0; for # [ #FOV
where d1 is the distance between transmitter and reflective point, d2 is the distance
between reflected point and receiver, q is the reflectance coefficient, dAw is a small
reflective area on the wall, a is the angle of incidence from the transmitter and b is the
angle of irradiance from a reflected point. The multipath character of the optical link
can be described by the light error propagation profile.
The tRTS and tCTS is the duration time of loss frame while transmission and tSIFS is
specify depletion time of SIFS. Therefore, we can analyze the COCC time and
understand the current packet transmission situation in optical wireless media access
layer once error propagation occurred.
the stream cypher to embedded an authentication function whether the data is modified
during data transmission. As information provided by the EU eSTREAM project, it is
proposed the WG-8 encryption algorithms that conform to have low-power con-
sumption encryption algorithm. These low-power consumption encryption algorithms
included MICKEY, WG-8 [20], Grain, and Trivium. For the best confidentiality, WG-8
has the lowest power consumption compared with rest of three encryption algorithms.
Once the LFSR is loaded with the key and IV, the apparatus runs for 40 clock
cycles. During each clock cycle, the 8-bit internal state S19 passes through the non-
linear WG-8 permutation with decimation d = 19 (i.e., the WGP-8(x19) module) and
the output is used as the feedback to update the internal state of the LFSR. The LFSR
update follows the recursive relation:
Sk þ 20 ¼ ðx Sk Þ Sk þ 1 Sk þ 2 Sk þ 3 Sk þ 4
ð5Þ
Sk þ 7 Sk þ 8 Sk þ 9 WGP-8(S19
k þ 19 Þ; 0 k 40:
After the key/IV initialization phase, the stream cipher WG-8 goes into the running
phase and 1-bit keystream is generated after each clock cycle.
While the WGP-8(x19) module permutes elements over F82 , the Tr() module com-
presses an 8-bit input to 1-bit keystream.
Sk þ 20 ¼ ðx Sk Þ Sk þ 1 Sk þ 2 Sk þ 3 Sk þ 4
ð6Þ
Sk þ 7 Sk þ 8 Sk þ 9 ; k 40:
4 Evaluation
In this study, we utilized Raspberry Pi 3B+ board as the experimental test bed. The
Raspberry Pi 3 Model B+ boards have been implemented in many smart factories and
can be represented as a classical embedded system test-bed. In our experiments, a
Raspberry Pi 3+ board was used as the master controller. The master controller is able
to decrypt the data from the cloud server and encrypt the data before sending them to
the cloud server. In our implementations, the experimental results showed that when
both the encryption and data communication functions are enabled, the Raspberry Pi
3B+ CPU was not overrun and could work well. Since the Raspberry Pi 3B+ is not a
high performance computer, this implied that the proposed mechanism could be easily
implemented with low computing power machines. As shown in Fig. 3, the proposed
mechanism will be able to support this low computing power machine and only
increase the total Raspberry Pi 3B+ CPU usage up to 49%. Moreover, once the data
transmission function enabled, the CPU utilization will reached 98%. Even both
encryption and data communication functions are enabled, the CPU utilization will still
running in the reasonable range.
488 S.-H. Chang et al.
5 Conclusion
In recent years, indoor positioning has draw lots of research attention. Due to the
wireless communication normally have well-known delivery problems, such as signal
fading; multipath propagation, signal obscured and interference problems, and these
will affect the indoor positioning accuracy. Accurate visible light data communication
utilizing received signal strength (RSS) approach that is the use of the signal from the
receiver with respect to the strength of the relationship between the distance attenua-
tion. This approach also can be applied to measurement of optical power absorption
and scattering across an air gap that is to be exposed by harsh factory environment. The
presented paper focus on a model-based understanding of particulates scattering
established, subsequent studies analyzed the impact of visible light scattering on smoke
air and the potential role of dust accumulation. However, future challenges will also
need to be highlight that are associated with visible light scattering measurements of
particulates in the presence of real-world surfaces.
References
1. Zurawski, R.: Industrial communication systems. In: The Industrial Information Technology
Handbook, Sec. 3, pp. 37–47. CRC Press, Boca Raton (2005)
2. Moyne, J.R., Tilbury, D.M.: The emergence of industrial control networks for manufacturing
control diagnostics and safety data. Proc. IEEE 95(1), 29–47 (2007)
3. Benzi, F., Buja, G., Felser, M.: Communication architectures for electrical drives. IEEE
Trans. Ind. Informat. 1(1), 47–53 (2005)
4. HART Field Communication Protocol Specification: HART Communication Foundation
Std., version 7.4 (2007). http://www.hartcomm.org/
5. Vitturi, S., Peretti, L., Seno, L., Zigliotto, M., Zunino, C.: Real-time Ethernet networks for
motion control. Comput. Stand. Interfaces 33(5), 465–476 (2011)
6. Willig, A.: Recent and emerging topics in wireless industrial communications: a selection.
IEEE Trans. Ind. Informat. 4(2), 102–124 (2008)
7. Wireless Systems for Industrial Automation, Process Control and Related Applications,
International Society of Automation (ISA) Standard ISA-100.11a (2009)
A Light Weight Stream Cypher Mechanism for Visible Light Communication 489
8. Sauter, T.: The continuing evolution of integration in factory automation. Proc. IEEE Ind.
Electron. Mag. 1(1), 10–19 (2007)
9. Sauter, T., Lobashov, M.: How to access factory floor information using Internet
technologies and gateways. Proc. IEEE Trans. Ind. Informat. 7(4), 699–712 (2011)
10. Gubbi, J., Buyya, R., Marusic, S., Palaniswami, M.: Internet of Things (IoTs): a vision,
architectural elements, and future directions. Future Gener. Comput. Syst. 29(7), 1645–1660
(2013)
11. Visible Light Communications Consortium. http://www.vlcc.net/
12. IEEE 802.15 WPAN™ Task Group 7 (TG7) Visible Light Communication. http://www.
ieee802.org/15/pub/TG7.html
13. O’Brien, D., Minh, H.L., Zeng, L., Faulkner, G., Lee, K., Jung, D., Oh, Y., Won, E.T.:
Indoor visible light communications: challenges and prospects. In: Proceedings SPIE, vol.
7091 (2008)
14. Pure VLC, Visible Light Communication: An introductory guide (2012). www.purevlc.net
15. Rajagopal, S., Roberts, R.D., Lim, S.-K.: IEEE 802.15.7 visible light communication:
modulation schemes and dimming support. IEEE Commun. Mag. 50(3), 72–82 (2012)
16. Adamko, L., Vojvoda, M., Jokay, M.: Statistical analysis of ECRYPT eSTREAM Phase3
Ciphers, EE časopis pre elektrotechniku a energetiku, pp. 193–196 (2008)
17. Antal, E., Zajac, P.: Analysis of the Fialka M-125 cipher-machine. In: Cryptologia (2013)
18. Engels, D., Fan, X., Gong, G., Hu, H., Smith, E.M.: Hummingbird: ultra-lightweight
cryptography for resource-constrained devices. In: 1st International Workshop on Light-
weight Cryptography for Resource-Constrained Devices. Tenerife, Canary Islands, Spain,
January 2010
19. Rukhin, A., et al.: Sp 800-22 Rev. 1a. a Statistical Test Suite for Random and Pseudorandom
Number Generators for Cryptographic Applications. Technical report, Gaithersburg, MD,
United States (2010)
20. Grosek, O., Vojvoda, M., Zajac, P.: Classical Ciphers. STU (2007). (in Slovak)
21. Perera, T., Hamer, D.: General introduction: Russian cold war era M-125 and M125-3MN
Fialka cipher machines (2005). http://enigmamuseum.com/mfialka.htm
22. Reuvers, P., Simons, M.: Fialka M-125: Detailed Description of the Russian Fialka Cipher
Machines. PAHJ Reuvers & MJH Simons (2009)
23. Robshaw, M., Billet, O. (eds.): New Stream Cipher Designs - The eSTREAM Finalists.
LNCS, vol. 4986. Springer, Heidelberg (2008)
24. Nawaz, Y., Gong, G.: WG: a family of stream ciphers with designed randomness properties.
Inf. Sci. 178(7), 1903–1916 (2008)
The 10th International Workshop on
Disaster and Emergency Information
Network Systems (IWDENS-2019)
Delay Tolerant Networks with Static Body
Object Detections by Mobile Sensors
for Disaster Information System
Abstract. Although the life safety information services are widely known if
there would be large-scale disasters, there have been some problems to use them
by people who are not good to operate or people who are injured. Therefore, this
paper proposes the static body object detection methods with the DTN for life
safety information system. In the system, the sensors on smartphones auto-
matically detect the abnormal static situations by the time and position differ-
ences based with the Markov Chain Monte Carlo methods, and the emergent
messages are automatically transmitted with the message priorities. Moreover,
those messages are transmitted to the servers by the Data Triage Methods with
the priority ID from the static body detections. Then, the experiments by the city
scale are discussed for the effectiveness of the proposed systems.
1 Introduction
It is now widely known that one can use the life safety information systems by PCs or
smartphones if disasters would occur. For example, the Google Person Finder [1] one
of the well-known services when the one needs the information of the life safety if one
cannot make a contact the relatives after large-scale disasters. Besides, there are cur-
rently many smartphone’s carriers that provide the life information service after the
Great East Japan Earthquake in 2011 [2]. Since the networks would not be available
after large-scale disasters, it is considered that these services are very effective.
However, it is also pointed out some problems of the systems. One of the problems
is that the users need to be good to operate smartphones or PCs. When there is a
disaster, the older people or children tend to be injured according to the previous
earthquakes. Also, if one gets heavily injured by the disaster, the one cannot send the
life safety information by using smartphones or PCs.
Therefore, it is proposed that the Delay Tolerant Networks with Static Body Object
Detections by Mobile Sensors for Disaster Information System [3] in this paper, and
this paper especially reported the simulation results by using the Map of Miyako, Iwate,
Japan. In details, the proposed methods firstly observe the gyro sensors and the
locational data such as GPS or WiFi periodically, and these data are applied for the
calculation based on the Markov Chain Monte Carlo (MCMC) methods [4] in order to
detect the abnormal static body detections such as injured by disasters. Then, if the
static body is detected, smartphones automatically send the emergent messages by the
DTN (Delay Tolerant Networking) [5] with the Data Triage Methods [6, 7].
In the followings, Sect. 2 describes the network architecture of the assumed life
safety information systems, and Sect. 3 presents the static body object detection
methods and the DTN routing with the priority ID from the static body detections in the
proposed methods. Then, the implementations of the simulator and the experiments are
shown in Sect. 4, and the conclusion and future study are discussed in Sect. 5.
2 Network Architecture
Figure 1 shows the assumed network architecture for the life safety information system
in this research. The network consists of the DTN networks in the damaged areas, the
gateway functions between the DTN and IP networks, and the IP networks in the safety
areas.
Fig. 1. The assumed system architectures of the life safety information system.
3 Proposed Methods
Fig. 2. Abnormal static body detections by the time threshold and the movement threshold in
this paper.
Therefore, the static body is detected by the abnormal static state in this research as
shown in Fig. 2. The time threshold and the movement threshold are calculated the
MCMS methods by the periodical observations from the gyro and the locational
sensors on the smartphones, and so these thresholds are optimally changed even if one
moves to the different environments. In the MCMS calculations [4], it approximates the
value function V(s) and the motivation function Q(s, a) by the reward ri as the formula
(1) and formula (2), where a is the learning rate ð0\a\1Þ.
X
1
Ri ¼ ri þ 1 þ cri þ 2 þ c2 ri þ 3 þ . . . ¼ cr r i þ 1 þ c ð3Þ
r¼0
496 N. Uchida et al.
The MCMS based calculations are held for the setting of the movement threshold
(Rth) and the time threshold (Tth) in the proposed methods for considering the different
environments in the injured victims.
Then, the proposed methods also apply the Data Triage Methods with the priority
ID from the static detections for the DTN routing. The DTN routing [5] is one of the
“store-and-forward” protocols, and so the node stores the received data when there is
no transmittable node around until the transmittable node appeared. However, it is
widely known that there are significant problems such as the large delay or latency, and
so there are various enhanced DTN approaches that are introduced by the Spray and
Wait [8], the Maxprop [9], and the Prophet [10]. The Data Triage Methods for the DTN
routing is also one of the enhanced DTN methods especially considered with the usage
under the disasters. In the Data Triage Methods [6, 7], the messages are once stored in
the queues in the mobile devices as shown in Fig. 3, and the messages are sorted by
order of the proposed priority ID. Here, the priority ID consists of the three digits of the
priority value from the static body detection, the 14 digits of the transmitted data, and
the 64 digits of the Android ID in the transmitted devices. Then, the upper data in the
figure is priority transmitted to the other transmittable nodes by the LTE D2D or the
Android WiFi P2P methods [3].
Fig. 3. Proposed Data Triage Methods with the priority ID from the static body detections
4 Experiments
In the previous paper [3], it was reported that the prototype system is implemented by
Nexus 5X (2 GB MEM, 16 GB Storage, Android OS 6.0.1, IEEE802.11a/b/g/n/ac),
Java 8.0, and Android Studio 2.1.3. Also, the MCMS libraries are customized by the
bases on the DL4j [11] and the HYDRA [12] in the system, and the previous study [3]
indicates the lower latency and the higher delivery rates in the field experiments.
However, the experiments were held on the academic campus, and so this paper
newly evaluated by the simulation with the city map of Taro, Iwate in Japan [6, 7]. In
the evaluations, the map routes of Taro have applied for the ONE (The Opportunistic
Network Environment simulator) ver. 1.6.0 [13], the comparisons between the Spray
and Wait routing and the Maxprop in this paper. Figure 4 shows the implemented map
in the simulator by the proposed methods, and the distance from the place to create
messages to the place to receive the messages is about 5.8 km. In the simulations, the
Delay Tolerant Networks with Static Body Object Detections 497
number of vehicles is differentiated with 25, 50,100, 200, and each vehicle run by a
random walk from 0 km/h to 50 km/h. The radio interface is IEEE802.11b that is the
100 m of the wireless transmission range, and the messages from 0.5 to 1.0 MB are
randomly created by every 50 to 60 s.
The results of the experiments are shown in Figs. 5 and 6. First of all, Fig. 5 present
the results of the delivery rates with the Spray and Wait routing and the Maxprop and
the proposed methods. It is pointed out that the messages with the priority ID in the
proposed methods were completely delivered with 100 vehicles, and that it was the
highest delivery rate than the others. On the other hands, the messages without the
priority ID were lower than the Epidemic and the Maxprop routing. Therefore, it is
supposed that the proposed methods were properly worked to the whole messages, and
that the significant messages were transmitted to the others faster than the non-priority
messages.
Secondly, Fig. 6 shows the results of the latency in the experiments. As the same
with the results of the delivery rates, the priority messages in the proposed methods
were lower latency than the others. On the other hands, the messages without the
priority ID were higher than the Epidemic and the Maxprop routing, but it is supposed
that the proposed methods were properly worked even in these experiments.
498 N. Uchida et al.
Fig. 5. Results of the delivery rate in the simulations by comparison with the Epidemic, the
Spray and Wait, the Maxprop, and the proposed routing methods.
Fig. 6. Results of the latency in the simulations by comparison with the Epidemic, the Spray
and Wait, the Maxprop, and the proposed routing methods.
As the results, it is considered that the proposed methods are effective even by the
city scales, and it might be helpful for the emergent purposes such as the rescues.
Delay Tolerant Networks with Static Body Object Detections 499
Although the life safety information services are widely known when there would be
large-scale disasters, it is significant to use them by people who are not good to operate
or people who are injured. Therefore, this paper proposes the static body object
detection methods with the DTN for life safety information system.
In the systems, the static body detections are held by the sensors on the smartphone
with MCMC algorithm, and the messages are transmitted to others by the Data Triage
Methods with the priority ID.
As the results of the simulation by the city scales, it is considered that the proposed
methods are effective for the assumed life safety information systems since the methods
are lower latency and higher delivery rate than these of other DTN routings. For the
future works, the field experiments in the evacuation drills by the prototype system are
planning, and the additional experiments of the static body detections are also planning.
Acknowledgement. This paper is the extend version of “Proposal of Static Body Object
Detection Methods with the DTN Routing for Life Safety Information Systems” in the 32nd
International Conference on Advanced Information Networking and Applications Workshops
(WAINA2018), May. 2018. Also, this work was supported by SCOPE (Strategic Information and
Communications R&D Promotion Programme) Grant Number 181502003 by Ministry of
Internal Affairs and Communications in Japan.
References
1. Google: Google Person Finder. https://www.google.org/personfinder/global/home.html.
Accessed Jan 2018
2. Shibata, Y., Uchida, N., Shiratori, N.: Analysis of and proposal for a disaster information
network from experience of the great East Japan earthquake. IEEE Commun. Mag. 52, 44–
50 (2014)
3. Uchida, N., Shingai, T., Shigetome, T., Ishida, T., Shibata, Y.: Proposal of static body object
detection methods with the DTN routing for life safety information systems. In: The 32nd
International Conference on Advanced Information Networking and Applications Work-
shops (WAINA2018), pp. 112–117 (2018)
4. Andrieu, C., Freitas, N.D., Doucet, A., Jordan, M.I.: An Introduction to MCMC for Machine
Learning. Kluwer Academic Publishers, Boston (2003)
5. Fall, K., Hooke, A., Torgerson, L., Cerf, V., Durst, B., Scott, K.: Delay-tolerant networking:
an approach to interplanetary internet. IEEE Commun. Mag. 41(6), 128–136 (2003)
6. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Proposal of data triage methods for
disaster information network system based on delay tolerant networking. In: The 7th
International Conference on Broadband and Wireless Computing, Communication and
Applications (BWCCA2013), pp. 15–21 (2013)
7. Uchida, N., Kawamura, N., Sato, G., Shibata, Y.: Delay tolerant networking with data triage
method based on emergent user policies for disaster information networks. Mobile Inf. Syst.
10(4), 347–359 (2014)
500 N. Uchida et al.
8. Spyropoulos, T., Psounis, K., Raghavendra, C.S.: Spray and wait: an efficient routing scheme
for intermittently connected mobile networks. In: Proceedings of the 2005 ACM SIGCOMM
Workshop on Delay-Tolerant Networking, WDTN 2005, pp. 252–259 (2005)
9. Burgess, J., Gallagher, B., Jensen, D., Levine, B.: MaxProp: routing for vehicle-based
disruption-tolerant networks. In: Proceedings of the 25th IEEE International Conference on
Computer Communications, INFOCOM 2006, pp. 1–11 (2006)
10. Lindgren, A., Doria, A., Scheln, O.: Probabilistic routing in intermittently connected
networks. ACM SIGMOBILE Mobile Comput. Commun. Rev. 7(3), 19–20 (2003)
11. DL4 J: Markov Chain Monte Carlo and Machine Learning - Deeplearning4j. https://
deeplearning4j.org/markovchainmontecarlo. Accessed Feb 2018
12. Warnes, G.R.: HYDRA: a Java library for Markov Chain Monte Carlo. J. Stat. Softw. 7(4)
(2002).
13. The ONE. https://akeranen.github.io/the-one/. Accessed Jan 2019
A Basic Study on Emergency Communication
System for Disaster Using LPWA
1 Introduction
There are many natural disasters in Japan, and disasters such as the Osaka prefecture
northern quake, the western Japan torrential rainfall, and the Hokkaido Ballistic Eastern
Earthquake occurred even in 2018 alone. When large-scale disasters occur, there is a
possibility that traffic is blocked due to damage to road structures, or communication
infrastructure is damaged, and communication is interrupted. Moreover, there is a high
possibility that power outages will last for a long time due to the damage of trans-
mission lines and others. Especially in the mountain area, the road is cut off due to cliff
collapse, etc., compared to the plain area, the isolation of the village, the possibility of
communication disruption and power outage is high. A study [1] is being conducted to
utilize amateur radio as a measure against isolation between mountainous areas. In
recent years, LPWA has attracted attention as a technology capable of communicating
over a wide range with low power consumption. In this research, we consider LPWA to
be used as a means of emergency communication when the communication infras-
tructure in the mountain area cannot be used due to disaster. In this research,
we propose to use LPWA which is more restrictive than the usual Internet commu-
nication in mountainous areas. We examined a method to utilize it for emergency
communication.
LPWA is an acronym for Low Power Wide Area. There are Bluetooth, ZigBee and so
on such as the art of low-power communication. LPWA can run more low-power those
arts and provides long-range communication. LPWA is ultra-narrow band, but LPWA
attracts a remark for IoT/M2M device, which is expected becoming popular, con-
necting with network. The main installation specifications of LPWA include LoRa,
SIGFOX, NB-IoT, and so on. In Japan, the usage with LoRaWAN method which is
used LoRa increases [2].
Generally, we can think that it uses LPWA as the means of collecting each sensor
information as the utilization for disaster prevention of LPWA. There is an experiment
of the usage in LPWA in the acquisition of the data of the level-gage of the underline
and the checker of the exposure-to-water by the inundation inside a levee in the
canalization. There is an experiment of the outgoing of the information on the disaster
prevention administration wireless which utilized LPWA [3].
In this study, we think of utilizing LPWA for an emergency communication in the
isolated mountainous area. We need a license for usage of the amateur radio, but
LPWA can be used by without a license. When using as means of communication in
self-, mutual help in the disaster, LPWA become means of communication which
anyone can use because the license is unnecessary.
The mountainous area is different from the plains and doesn’t often put on by the
direct-communication, foreseeing the end of the communication because there are
mountains as blocker. But there is possibility that the mountain reflection or mountain
diffraction can be used for the radio-propagation. When installing an antenna in the
heights in the plains, it is necessary to build the steel tower and so on. But when
installing an antenna in a summit or a hillside in mountainous area, we can wireless
communication in the heights.
In spite of short distance communication, it is difficult to communicate with direct-
communication in Low-Low place wireless communication like Terminal A-Terminal
B in Fig. 1. Because of installing a wireless repeater in a summit or a hillside, we can
A Basic Study on Emergency Communication System for Disaster Using LPWA 503
Repeater
Our own development programs are divided into three programs, for sending,
repeating, and receiving. And for sending program is divided in two programs for
sending free message and for sending auto increment number for communication test.
For repeating program sends the just receiving message. For repeating and sending
programs have sender information in header on the message. And we can see the
message and judge from whether repeater or terminal using for receiving program. In
addition, we development reading receive message function by using Open JTalk [5]
for using in driving a car. These programs are output logs (CSV) about each
send/receive message and time, RSSI and so on.
Noson park. And the state of the sending side (Akiba Dam) at the time of communi-
cation experiment is shown in Fig. 3 (right).
Fig. 3. The location of LPWA communication experiment in mountainous area (left), and state
of sending side (Akiba Dam)
Fig. 4. State of sending side (Shimohirayama Noson Park) (left), and state of receiving side
(car) (right)
506 H. Yuze and S. Nabeta
On January 6th to 7th, we have grasped the range of communication from high
altitude prospects. At that time, based on Shimohirayama Noson Park, we installed
LPWA device in the car and confirmed communicability while driving car. Figure 4
(left) shows a state of Shimohirayama Noson Park as a sending side. Figure 4 (right)
shows a state of car as a receiving side.
Figure 5 shows the locus of movement of the car moving from Shimohirayama
Noson Park for reception. It is the point where the red dot in the figure was recorded as
the GPS logs. In Fig. 5 the big dots were receivable places, the small dots were not. As
a result of receiving experiments while driving by car, it was confirmed that both Akiba
Dam and Hamamatsu Tatsuyama Forest and Culture Center can communicate from
Shimohirayama Noson Park.
Dam with PC, send it again at the relay device at Shimohirayama Noson Park, and
receive the relayed message at the Hamamatsu Tatsuyama Forest and Culture Center.
The received message is displayed on the PC on the receiving side, and the PC
synthesizes the message and reads it. By sending several patterns of messages, we
confirmed that we can send messages without problems between Akiba Dam and
Hamamatsu Tatsuyama Forest and Culture Center. After that, we further changed the
reception location and confirmed how far the message communication can be done.
Figure 6 shows the message displayed on the receiving computer.
University of Shizuoka to which the authors belong is located at the foot of Nihondaira,
the altitude is about 100 m, and it is located in a well-prospective place. It is also the
self evacuation area and fully autonomous anti-disaster base station [6, 7] have also
been established. We conducted a communication experiment between university and
Shizuoka prefecture disaster hospital with a wireless network. Therefore, we conducted
a communication experiment on the actual communication range of the LPWA device
on September 17th, 2018.
In the University of Shizuoka as the sending side, we set up a PC that launched the
LPWA device and the sending program which is only sent auto increment number and
in the car as the receiving side, we installed the LPWA device. We carried out reception
experiment while moving by car. As a result, the maximum coverage area in this
experiment was near the Surugawan-Numazu Smart Interchange on the Shin Tomei
Expressway, which is a straight distance of 37 km from University of Shizuoka. This
result is shown in Fig. 7. The map shown in Fig. 7 is a marker of each site added to the
508 H. Yuze and S. Nabeta
electronic map (tile) from Geospatial Information Authority of Japan. In Fig. 7, “●”
indicates University of Shizuoka and “★” indicates the maximum communicable point
in this experiment.
8 Conclusion
References
1. Ueno, K., Mori, A., Nakano, S., Yoshida, A.: Utilization of amateur radio in isolated
mountainous village due to disaster. In: Proceedings of the JSCE Earthquake Engineering
Symposium, vol. 30 (2009). 4-0036
2. Tanaka, M.: Recent development of LoRaWAN in Japan, IEICE Technical report, vol. 117,
no. 297, SRW2017-58, pp. 59–64 (2017)
3. Yuze, H., Shibata, Y., Sugawara, K., Nakagawa, Y., Ono, M.: Communication experiment on
information dissemination by disaster prevention administrative radio with wireless
communication technology “LPWA”. In: Proceedings of the 19th Conference, Japan Society
for Information Studies, pp. 120–121 (2017)
4. Green House: Development board, RM-EVSA-92AS-2. https://www.green-house.co.jp/
products/rm-evsa-92as-2/
5. Open JTalk: Open Jtalk. http://open-jtalk.sourceforge.net/
6. Yuze, H., Nabeta, S., Ito, Y.: Development of autonomous anti-disaster base station by
combination of satellite and wireless networks. In: 2015 Workshops of 29th International
Conference on Advanced Information Networking and Applications (WAINA), pp. 524–528
(2015)
7. Yuze, H., Nabeta, S., Ito, Y., Kosaka, H., Shibata, Y.: Improvement of autonomous anti-
disaster base station by combination of satellite and wireless networks. In: 2016 Workshops of
30th International Conference on Advanced Information Networking and Applications
(WAINA), pp. 667–672 (2016)
Platform System Based on LoRa Mesh
Network Technology
Abstract. Although the LoRa private network has started to be used in various
applications due to its characteristics, there are only arrangements for carrier
sense time, transmittable time, and post-transmission standby time on the LoRa
standard and Radio Law, It is necessary for the user side to independently deal
with behaviors related to retransmission of data lost at the time of collision and
error correction function. In addition, since it is a relatively new wireless
communication method, it can be said that it is almost in a disorderly state
without unified guidelines and tacit agreements that have been accumulated in
the usual way. This situation is particularly problematic when various services
start utilizing LoRa with their own protocols, so a uniformly available LoRa
platform is necessary.
2 Methods
In order to exchange and share information with neighboring shoulder system in real
time, we will integrate N different wavelengths of different wavelengths and develop N
wavelength cognitive radio. Then, the radio link that maximizes the evaluation function
based on the necessary communication quality (throughput, delay, packet loss rate,
RSSI) is determined according to the change of the communication environment
(communication distance, noise ratio, interference, etc.). By maximizing communica-
tion distance and total data transfer amount by realizing inter-vehicle communication.
For this reason, we develop a function that can select an optimal wireless link from N
wavelength radio links by SDN (Software Defined Network) protocol function. As the
N wavelength radio, a radio for IoT such as a wireless LAN or LoRa radio corre-
sponding to the IEEE802.11n (2.4 GHz)/ac (5.6 GHz)/ad/60 GHz) standard that is
available for license generation and is the next generation wireless network We develop
an N wavelength cognitive radio unit by suitable combination from standards etc. In
particular, LoRa wireless communication uses broadcast communication, if collision of
wireless signals occurs due to simultaneous transmission, retransmission is not per-
formed. Also, in consideration of various communication environments, nodes that
perform communication have hidden terminal problems because radio wave reach
ranges are extremely different. In order to avoid simultaneous transmission, LoRa mesh
network technology implements time-division access control method after time syn-
chronization by GPS. It is assumed that all the radio stations operated in the system are
numbered by ID and the total number is also grasped. In this premise, based on the ID
dynamically allocate time slots that it can transmit. For example, assuming that the
transmission time per wireless device is 10 s (considering the time required for backoff
to avoid collisions with other systems), when using five wireless stations, these five
devices do not have any problems A time longer than 50 s that can be transmitted is
assigned as a transmission cycle. In this transmission cycle, as shown in Fig. 1,
transmission time slots are assigned based on respective IDs. As a result, occurrence of
collision in communication in the own system is avoided. The amount of data that can
be transmitted per unit time of the wireless device and the communicable distance vary
depending on factors such as the spreading factor of LoRa, the used channel band-
width, the surrounding communication environment, and the like. Also, since the
transmission cycle varies according to the number of radio devices used in the system,
slot design of the entire system is required in advance according to the requirements.
Therefore, in this research, by appropriately designing the spreading factor of the
LoRa, the used channel bandwidth, and the transmission cycle according to the fol-
lowing three kinds of application systems on the basis of the LoRa mesh network
technology, Verify feasibility. By the way, data relaying is done by using flooding.
However, in addition to the mechanism to avoid transmitting the same information
multiple times, in order to reduce the number of flooding transmission, the following
implementation is carried out.
First, by sending information listing the sender IDs of the signals received by each
node in the past as control signals at regular intervals, it is possible to grasp the range
within which their neighboring nodes can communicate. When a newly received
512 G. Sato et al.
Fig. 1. Relationship between transmission cycle time and transmission time per wireless device
information is received, if there is a node that advertised the information earlier before
its own transmission timeslot comes around, if the neighboring node that it covers is
adjacent to If it is all covered, it will not retransmit. As a result, retransmission of
unnecessary information is suppressed.
Fig. 3. Variable slot overview for collecting sensor data with different data amount
a, b, c are the weight of moving average. m is the measurement results for each
time.
Then, smoothing values are normalized by this formula.
8 ni li
>
< ð1 ui li Þ 10 ðli \ni \ui Þ
Si ¼ 1 ðni ui Þ
>
:
9 ðni li Þ
ni l i
Si ¼ ð Þ 10
umax li
514 G. Sato et al.
n is the smoothed value, l is the minimum value, and u is the maximum value in
monitoring history. Finally, it will be evaluated using the normalized value by this
formula.
x, y, z are the weight of the network policy. Basically, use the link with the best Pi
value. However, since there is a characteristic that the throughput is very low with
respect to the LoRa radio, basically it is used like a control channel for exchanging
information for the above judgment and key information for connection.
For field demonstration experiment from next fiscal year, we are currently studying
outdoor prototype system as below (Fig. 4).
4 Summary
In this paper, we propose a unified LoRa platform that does not rely on existing
communication infrastructure, and designed to adapt to time-division access control
method necessary for realizing it. In addition, we developed a system for sharing
vehicle position information using a self-operated network using a basic time division
access control method, actually installed a radio at the hospital rooftop of the city
center, and then sent out-of-line communication We verified whether or not desired
communication can be done even in environmentalization. As a result, it was confirmed
that even in the downtown area, it is sufficiently possible to communicate about the
sharing of the position information of the vehicle within a radius of about 2 km. In
addition, it is shown that time sharing access control communication using basic
technology of LoRa platform enables simultaneous sharing of information among
multiple sites.
Reference
1. Inoue, M., Ohnishi, M., Peng, C., Li, R., Owada, Y.: NerveNet: a regional platform network
for context-aware services with sensors and actuators. IEICE Trans. Commun. E94-B(3),
618–629 (2011)
Construction of a Disaster Response Automatic
Extraction Support System
1 System Configuration
In order to solve the issues of previous research [1, 2], we proposal the disaster
response automatic extraction support system. The configuration of the system con-
structed in this research is shown in Fig. 1. This system supports decision-making of
municipal disaster countermeasures headquarters installed in each local government at
the occurrence of a disaster. In normal times, the user accumulates past disaster
response records and local disaster prevention plans in this system. On the other hand,
in emergency times, the user searches for disaster response records and shares infor-
mation by using the accumulated data. Also, the user visualizes registered disaster
information and social media information related to the disaster. In this research, we
refer to the “disaster information accumulating system” for the system used at normal
times and the “disaster information visualization and analysis system” for the system
Fig. 1. The system configuration in the disaster response automatic extraction support system.
518 T. Ohyanagi et al.
The disaster information accumulating system and the disaster information visu-
alization and analysis system communicate with the Disaster Historical Records
Visualization and Analysis System API (DHRVAS API) on the server for data man-
agement and visualization. By communicating with the DHRVAS API, this system
realizes reduction of the amount of data received. In addition, this system realizes
improvement of the response speed of the system and reduction of communication
load. In order to realize this function, we use asynchronous communication technology
and standard data format. We adopted Fetch API provided by JavaScript in asyn-
chronous communication technology and JSON (JavaScript Object Notation) in stan-
dard data format. In addition, we have increased system extensibility and diversion by
SOA (Service Oriented Architecture) which clearly divides server application and
client application. For example, when adding a new function, it is possible to add the
logic to the server and use the function by simply calling the API on the client side.
Moreover, by calling the API, it is easy to facilitate data linkage with other systems.
This also facilitates the introduction of new systems. The method that the existing
system uses the data of this system via the API is shown in Fig. 2.
This system uses Relational Database (RDB) and NoSQL Database (NoSQL DB)
as database server. The RDB consists of disaster classification table, disaster prevention
plan table, multimedia table, disaster table, disaster response record table, disaster
response phase table, region table, prefecture table, municipal table, user table, and
disaster category table. The ER diagram of each table is shown in Fig. 3.
Construction of a Disaster Response Automatic Extraction 519
In this research, various kinds of disaster information such as disaster name, disaster
response record, local disaster prevention plan etc. are accumulated in the database.
This will promptly provide past disaster information to users in emergency times and
support disaster response decision making. In order for the user to extract past disaster
information, data is retrieved by using parameters such as date and time, disaster
classification, disaster response phase, and the like. Furthermore, this system provides a
full-text retrieving function that extracts texts most approximate to the input search
keywords. Full-text retrieve is one of the document retrieval methods and is different
from general retrieval methods (progressive scan retrieve and sequential retrieve) for
retrieving documents one by one. By scanning all documents in advance and con-
structing index data, high-speed retrieve is realized. The outline of full-text retrieve is
shown in Fig. 4.
The disaster response automatic extraction support system consists of the disaster
information accumulating system and the disaster information visualization and anal-
ysis system.
Next, the logout function is displayed by selecting the menu button installed at the
top of the main screen of the system with the user logged in (Fig. 6). When the logout
button is pressed, the user information and all the information cached by the system are
deleted.
Construction of a Disaster Response Automatic Extraction 521
Fig. 6. Logout screen of the disaster information accumulating system and the disaster
information visualization and analysis system
Fig. 7. Basic information selection screen of the disaster information accumulating system
visualization and analysis system. The basic information selection screen of the disaster
information visualization and analysis system is shown in Fig. 8. On the basic infor-
mation selection screen of the disaster information visualization and analysis system,
the user selects “disaster response phase” and “season” in addition to “disaster type”
and “local government name” which were installed in the disaster information accu-
mulating system. The season is automatically calculated from the current date and time
by the system. After completing the data input, pressing the confirm button transits to
the next screen. At this time, like the disaster information accumulating system, the
information entered on this screen and the user’s authentication information are stored
in the local storage of the Web browser.
Fig. 8. Basic information selection screen of the disaster information visualization and analysis
system
4 Conclusion
In this research, we conducted a hearing survey on local governments and clarified the
current disaster response process. As a result, we found that the municipal disaster
countermeasures headquarters is spending a great deal of time collecting and orga-
nizing damage information and shelter information. This hinders prompt decision
making within the municipal disaster countermeasures headquarters. Therefore, in this
research, we extracted requirements for smoothly making decisions from the disaster
response process obtained in the hearing survey. Based on this requirement, we have
constructed the disaster response automatic extraction support system aiming to sup-
port decision making of municipal disaster countermeasures headquarters.
References
1. Takahagi, K., Ishida, T., Sakuraba, A., Sugita, K., Uchida, N., Shibata, Y.: Construction of a
mega disaster crisis management system. J. Internet Serv. Inf. Secur. (JISIS) 5(4), 20–40
(2015)
2. Ishida, T., Hirohara, Y., Uchida, N., Shibata, Y.: Implementation of an integrated disaster
information cloud system for disaster control. J. Internet Serv. Inf. Secur. (JISIS) 7(4), 1–20
(2017)
Network Performance Evaluation to Realize
N-Wavelength V2X Cognitive Wireless
Communication System
1 Introduction
Intelligent Transport Systems (ITS) for road transport realizes safer driving which is
under consideration of road traffic or road environmental information. ITS allows to
perform road traffic surveillance with roadside unit (RSU) which has several types of
sensor such as temperature or CCTV. These collected information is realized for road
administrative operation and make road traffic safely providing information for drivers.
Another evolution of ITS can be identified for vehicle onboard sensor gets popu-
larly equipment. Some modern commercial available vehicles provide notification of
forward traffic signs for drivers, for example, when the vehicle is closing to stop sign at
the intersection, vehicle recognizes it using onboard camera and make notification for
driver. Onboard sensor allows to realize lane departure warning or collision avoidance
functions, this function is categorized into SAE’s 1 or 2 autonomy level [1]. In dis-
cussion for realization of higher level of autonomy vehicle, there is a considerable
challenge how the vehicle should collect forward road surface condition. Any auton-
omy vehicles should be considered of maneuvering which corresponds to road
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 524–536, 2019.
https://doi.org/10.1007/978-3-030-15035-8_50
Network Performance Evaluation 525
condition in order to safely control even at snowy or icy road. Especially, in the cold
district, road condition often changes into short distance and in short time which is
caused by weather changing, thus it is required to collect actual road surface condition
in consecutive analyzing on sensing vehicle.
Due to above-mentioned methods which are based on onboard analyzing, the
required system functions must be supposed for system functions supposed for use of
wireless communication in effort to exchange with other vehicles. Floating car data
(FCD) is a one of practical V2X communication technology which depends on public
cellular network, it has characteristics for being blind zone to originate from out of
service area such rural or mountainous area, thus, this system should design inde-
pendent from public cellular network.
Vehicle-to-Everything (V2X) communication is a concept of Vehicle-to-Vehicle
(V2V) communications, Vehicle-to-Road (V2R), and other vehicular related commu-
nications together and makes data communication without cellular or other public
wireless access network possible between vehicle and others or RSUs. As a back-
ground, modern ITS uses dedicated short-range communication (DSRC) as known as
IEEE 802.11p standard for delivering data. However there is limitation of transmission
speed and communication range which are not enough when road condition
exchanging system increases amount of delivering data or V2V communication.
Another key technology to realize a connected car is low-power wide area (LPWA)
networking which provides long-range distance in several kilometers wireless com-
munication but provides quite low throughput of data delivering at a time.
This paper proposes a wireless communication system designed for V2X envi-
ronment in order to deliver condition of roads. In the principal design, the V2X
communication deals with associated road state such dried/wet/snowed/icy/whiteout
etc., and its precise location. The decision of road states is performed onboard node into
each vehicle by collecting and calculating of multiple sensor value, afterwards the
proposed system sends the result to roadside unit or other vehicle directly. In this paper,
we introduce three different wavelength wireless devices which contain of 2.4 GHz,
5.6 GHz and 920 MHz LPWA wireless network. In our design, 2.4/5.6 GHz wireless
LAN (WLAN) bands conduct data transmission mainly. Authentication and initial-
ization of connection on WLAN bands connection both will be sent over 920 MHz
band before vehicle is in WLAN range. The system design is designed for very high
speed and highly efficiency wireless network bandwidth in road condition data trans-
mission even vehicle is moving fast and road information becomes richer. Beam-
forming technology which is implemented 802.11ac standard allows us to realize to
realize more efficiency delivery of packets and to reach farther devices.
We have evaluated network performance to confirm our approach by field exper-
iment. The result indicates several times of higher efficiency for data delivering which
is compared with existing implementation based on 2.4 GHz WLAN standard.
526 A. Sakuraba et al.
2 Related Works
Road sensing technique processes in order to determine what road condition is cor-
responded to by analyzing of sensor output which is placed on moving vehicle. These
methods had been introduced in field of pavement engineering with non-contact
onboard sensor to collect pavement condition while probing vehicle is running. In this
method, dynamics sensors especially acceleration sensor is one of popularly sensor to
obtain road condition such pothole or bump of pavement. Chen et al. described their
road monitoring system to process sensor value for power spectral density in order to
estimate the International Roughness Index (IRI) using 3-axis acceleration sensor. As
an example for ideas based on using of non-dynamics sensor for road condition
sensing, Nakashima et al. discussed a method which is implemented with ultrasonic
sensor. Observation of quasi electrostatic field (QEF) between road surface and vehicle
body based road sensing method brought suggestion which was useful to determine it,
although the result showed that it required more accuracy method for determination of
road surface condition is equivalent to compacted snow and icy.
Almost practical implementations in order to provide floating car data function is
based on public wireless network such 3G/3.5G/3.9G/4G cellular network which is
available to deliver data in from several hundred kbps to dozen Mbps throughput.
Several automotive industries provide them in the market. For example, Honda
Internavi [2] allows precisely estimation of traffic jam which is jointly analyzed with
information from RSU and floating data from other vehicle. In their architecture, V2V
communication depended on public cellular network. CarTel [3] has V2R communi-
cation on private wireless network which is short-range vehicular adhoc network
(VANET) based on 802.11b WLAN, it was possible to deliver data which is capable
about 30 KiB/s end-to-end throughput from vehicle to internet.
We can cite a number of VANET proposal methods which is based on DSRC or
802.11p, and 802.11a/b/g/n for the popular wireless standard using for V2R or V2V
communication device. WAVE system [4] is a prototype of DSRC operated on
5.8 GHz band and designed to notify traffic signs for driver with V2R communication.
RSUs which equips the improved DSRC based wireless device are now in operation at
actual expressway of Japan. Bellavista et al. proposed a V2X communication protocol
for traffic light state exchanging method among other cars and RSU [5]. Their method
considered to operate in realistic scenario and congested urban area, it designed on
2.4 GHz 802.11b wireless with group-based communication protocol.
V2X applications especially require wireless communication short latency and less
packet loss ratio not only high efficiency data delivering, there is several implemen-
tations of cognitive radio for V2X communication which consists of multiple wireless
link and switches it situationally. Ito et al. developed 2-wavelength wireless commu-
nication consists of 920 MHz and 802.11b/g/n 2.4 GHz WLAN device to reduce
latency which is caused time of authentication and association which is performed
before delivering on WLAN band, they confirmed that the system allows to increase
amount of data delivering between vehicle and RSU [6].
Network Performance Evaluation 527
This section describes road condition sharing process of our proposal system which can
obtain sensor value and determine road condition on the vehicle, and then exchanging
with other vehicle and/or RSU and share among neighbor vehicles.
3.1 Overview
Figure 1 illustrates an overview of our system. Our proposed system can share with
highly real-time road states for drivers in manual driving.
Onboard unit is loaded on each vehicle which consists of Sensor Server and SMB
(Smart Mobile Box). Sensor Server deals collecting of multiple onboard sensor output
firstly. SMB analyzes for decision road states of up to current travelling point which
describes as road state corresponding to dried/wet/snowy/icy/whiteout etc. with gran-
ularity. Next, SMB associates road state and their location. Then SMB delivers it with
raw sensor data to RSU called Smart Relay Shelter (SRS) on N-wavelength V2R
cognitive wireless network. On other side, roadside server provides road data and states
of wide area is collected by other vehicle via the same network. This cognitive network
also takes V2V communication directly with a vehicle which is heading to the opposite
direction from own vehicle. At the same time, SRS also perform to deliver regional
road states and acquires other area road surface states with cloudlet on cloud computing
resources, therefore our approach allows to provide service is public cellular network
independent.
Finally, the system provides visualization of road states based GIS (Geographic
Information System) for users of road such drivers or operation managers of public
transportation system such bus or taxi. For operation managers or driver who is not in
528 A. Sakuraba et al.
driving, our proposal system represents road states as generic GIS designed for using
on smart device and PC. Each vehicle equips road surface warning system in order to
represent and warn for driver who is in driving, if ahead road condition would get
worse which is described as icy, whiteout, etc. system will warn it in visually on
display device and playout of warning message by human voice.
Continually acquiring of road states information is required to provide precise
information, it could be efficient to load onboard unit on frequently running vehicles
along the same route such bus.
These functions allows acquiring, sharing, and representing of road condition, it
could realize safer road traffic even winter season.
Sensor server monitors output value from sensor group which is composed of
various sensor of acceleration, gyro, geomagnetic, temperature, humidity, atmospheric
pressure, quasi electrostatic field (QEF), RGB image, and positioning of vehicle in
order to determine road surface state where vehicle is running. Sensor server is based
on edge computing architecture and processes acquired sensor values to identify road
state by machine learning algorithm. After road state decision has been completed, the
result of analyze is sent to SMB.
SMB has several wireless networks with different wavelength wireless network
device and decision maker which wireless device should be delivered to other SMB or
SRS.
At this stage we designed our proposal system has two of different wavelength
network device but extensibility is considered use of N-wavelength wireless link, such
IEEE 802.11ax or 802.11ay standard based one, into system design. Detail of using
wireless networks are described in Sect. 3.3.
Network Performance Evaluation 529
(1) Message & data flow in V2R (2) Message & data flow in V2V
Fig. 3. Transmission of data and message flow. (1) describes V2R transmission in the event of
destination vehicle requests information to SRS. (2) describes destination vehicle requests it to
source vehicle in adhoc wireless communication
Messaging flow of V2R scenario is illustrated as Fig. 3-(1). Initially, SRS broad-
casts own existence information via 920 MHz band, which corresponds to beacon
frame of access point in 802.11 based WLAN. This information includes traffic
direction which SRS collects and provides road states. When running vehicle receives
broadcast message from SRS which is installed in the same traffic direction as the
vehicle is moving, SMB responses to SRS with own node ID. Responding to that, SRS
assigns an IP address for WLAN on SMB and then deliver WLAN network infor-
mation for SMB. These information is composed of assigned IP address, netmask,
default gateway address, ESSID of access point, WPA passphrase etc. due to SRS
requests SMB to connect Infrastructure Mode in V2R communication. SMB activates
WLAN device after connection information has been received and waits for vehicle is
in range of WLAN of SRS. When the vehicle entered into WLAN coverage area, SMB
attempts to take 4-way handshake for WLAN connection. If WPA authentication has
been succeeded, it also checks for continuity by ICMP message exchanging. At this
point WLAN connection has been established completely, thus both vehicle and SRS
are able to exchange road state information over TCP each other.
Figure 3-(2) denotes message flow which describes V2V communication from
initiation to be established. Message and data flow is very similar procedures, firstly
destination vehicle which requires road state information from other vehicle, waits
broadcasts own existence information from source vehicle which has already road state
information, via 920 MHz band. If SMB on destination vehicle received broadcast
Network Performance Evaluation 531
4 Prototype Environment
5 Evaluation
Fig. 4. Device connection and configuration of experimental setting on SRS and SMB
Fig. 5. Overview of both vehicle direction trials and installation of SMB wireless units
We had planned this evaluation in actual public two-lane road which is illustrated
in Fig. 6. The length of evaluation section on road was 1.1 km and road alignment was
almost straight and almost smooth terrain but there was several grade section which
includes 3.3% grade in 90 m length. We placed SRS on sidewalk where is located next
to the westbound lane. The radio antenna was fixed on an antenna pole and their height
Network Performance Evaluation 533
from the ground was 2.2 m for LPWA, 1.7 m for 802.11ac, and 1.4 m for 802.11b/g/n.
802.11ac WLAN unit on SRS faced to direction across of road in 90°.
The condition of road traffic density was medium and weather was snowfall.
5.4 Result
We have ran 5 trials each three different types configuration and both vehicle’s
directions. Figure 7 plots delivered data amount which is succeeded to send from SMB.
Generically, 802.11ac link can deliver more efficiently in every trials. The previous
method which uses 2.4 GHz link was confined to deliver averagely sent 53.1 MiB
(s.d. = 8.5) in eastbound and 52.8 MiB (s.d. = 7.1) in westbound trials, nevertheless
our system averagely delivered 218 MiB (s.d. = 109.7) from moving vehicle to SRS in
eastbound and 301 MiB (s.d. = 65.8) in westbound. The result indicates that our
method was 4.1 to 5.7 times faster than 802.11b/g/n based system.
We also measured delivered packet in environment which based on generic
802.11ac, the averaged delivered data was 125 MiB (s.d. = 175.2) in eastbound trials
and 259 MiB (s.d. = 160.0) westbound one. The result recorded three times of data
delivering failure in generic 802.11ac setup and it contains two times timeout while
FTP session establishing, and once authentication failure on WPA2 before delivering
IP packets.
5.5 Discussion
The result proved that high-speed wireless link even short distance coverage, such
802.11ac which is a sort of license-free bands, could be allowed to deliver a large
534 A. Sakuraba et al.
(1) Sent data amount to SRS during 5 trials when (2) Sent data amount to SRS during 5 trials when
vehicle moves to west vehicle moves to east
Fig. 7. Plotted delivered data amount while V2R data exchanging over each wireless system
amount data from moving vehicle to SRS. For example, it is capable to exchange long
time RGB sensor output which corresponds to in 10 min recording if averaged size of
compressed road surface image equals to 500 KiB and it is recorded in a frame per
second. We suggest reason of availability for large amount data delivering was caused
of beamforming technology on 802.11ac standard has an effect on high throughput
which was observed in our field experiment.
The effect for packet delivering performance improvement on 5.6 GHz band with
920 MHz LPWA is not able to be concluded in this time, due to unstable performance.
Thereby we are required to repeatedly examine both network performance in the field.
Another topic for discussion is performance difference which is depended on
relationship between SRS placed location and traveling lane of the vehicle was contact
with SRS or not. In our result, westbound trials demonstrated 38% better performance
which compared with eastbound trials for delivering packet. This result could be
attributed to multiple factors such as oncoming vehicle, objects around circumference
of road such building, trees, etc. Especially, we surmised the distance among onboard
WLAN units and SRS in westbound trials which was clearly shorter than eastbound.
This result suggests to install multiple WLAN units on vehicles such right, left, and rear
of the body of car, in effort to obtain stronger signal between SRS and SMB, therefore
it realize better network performance.
6 Conclusion
This work proposed a design for N-wavelength wireless communication system for
V2X environment to exchange road surface conditions. This system has three of
fundamental functions; observing of physically state of road surface by multiple sen-
sors on the vehicle, analyze these sensor output for decision of road surface conditions,
Network Performance Evaluation 535
and transmit calculated road condition to roadside device or other vehicle to share road
condition which the vehicle travelled until now. We designed that this road surface
state provides to warn drivers both own and others if ahead road surface condition will
be getting worse. Road condition is also provided in road service GIS for person who is
not a driver.
Our system uses various different wavelength wireless networks among vehicles
and RSUs which is not depended on public wireless WAN, in order to be available the
functions even mountainous district. Our current work is implementing with two-
wavelength wireless network consists of 920 MHz and 2.4/5.6 GHz license-free bands.
The system delivers road state information over 2.4 GHz, send raw sensor values over
5.6 GHz, and exchanges network or authentication information for WLAN over
920 MHz LPWA band, this design allows onboard cognitive wireless communication
unit to immediately send packet from WLAN devices when vehicle will be just
entering into WLAN coverage of SRS. This design is considered to improve packet
delivery efficiency for communication among moving vehicles.
We have measured V2R communication performance in actual road environment.
The result of evaluation proved that our proposed approach is 4.1 to 5.7 times faster
than previous approach which is based on 2.4 GHz WLAN. However effect on network
performance improvement using 920 MHz LPWA was not clear. The result also
suggested to install multiple WLAN unit on the each sides of vehicle.
In our future work, we are considering a new networking protocol to realize the
most efficiency on WLAN band which is able to packet delivering less than a second
when vehicle moved into range of WLAN. We are now investigating of combination of
LPWA and FILS (fast initial link setup) which is standardized as IEEE 802.11ai [7].
References
1. SAE International: Taxonomy and Definitions for Terms Related to Driving Automation
Systems for On-Road Motor Vehicles J3016_201806 (2018)
2. Honda Global, October 17, 2003 Honda Introduces World’s First Congestion Prediction
Function for it’s InterNavi Premium Club Car Navigation Service. https://global.honda/
newsroom/news/2003/4031017-navi-eng.html. Accessed 20 Jan 2019
3. Hull, B., et al.: CarTel: a distributed mobile sensor computing system. In: Proceedings of the
4th Conference on Embedded Networked Sensor Systems (2006)
4. Tsuboi, T., Yamada, J., Yamauchi, N., Hayashi, M.: Dual receiver communication system for
DSRC. In: Proceeding of 2008 2nd International Conference on Future Generation
Communication and Networking, pp. 459–464 (2008)
5. Bellavista, P., et al.: V2X protocols for low-penetration-rate and cooperative traffic
estimations. In: Proceedings of 2014 IEEE 80th Vehicular Technology Conference,
pp. 1–6 (2014)
536 A. Sakuraba et al.
6. Ito, K., Hashimoto, K., Shibata, Y.: V2X communication system for sharing road alert
information using cognitive network. In: Proceeding of 8th International Conference on
Awareness Science and Technology, pp. 533–538 (2017)
7. IEEE 802.11ai-2016 - IEEE Standard for Information technology - Telecommunications and
information exchange between systems - Local and metropolitan area networks - Specific
requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer
(PHY) Specifications Amendment 1: Fast Initial Link Setup (2016)
A IoT Based Disaster Information Platform
for Challenged Network Environment
in Snow Countries
Abstract. In many snow countries, since the traffic information and commu-
nication networks infrastructures are not well developed due to their inherent
weather and geological conditions and rapid progress of aging and depopulation,
the resident’s safe livings are always suffered from those challenged environ-
ments, such as disasters, heavy snow, iced road conditions. In order to release
from those serious problems, a new IoT based information platform is intro-
duced in this paper. On the proposed platform, mobility such as cars and drones
can performs information and communication means to organize an information
infrastructure in challenged network environment where both wired and wireless
network are not well installed such as mountain areas and coast areas in snow
countries. In the disaster cases as emergency situations, drones are launched
from cars and can organize temporal wireless mesh networks in the sky and can
cover the wide communication areas. Thus by combing both the cars and
drones, very important information related resident’s life and property can
quickly and flexibly exchanged using this platform. In this paper, the general
system configuration and architecture and its expected application for disasters
are discussed.
1 Introduction
Many areas in Japan is occupied by mountains and forests. However, in those areas,
aging population, depopulation and declining economic activity are rapidly advancing
[1, 2]. Particularly in the cold areas, the residents are forced to drive in the chronic
serious bad road conditions, such as snow storm, frozen road, snow slide, whiteout. For
this reason, the number of traffic accidents are increasing as the aging and abnormal
weather conditions are advancing. The information and communication environment
along the roads in local areas are not well developed compared with urban areas. Once
a traffic accidents or disaster occurred, information collection, transmission and sharing
are delayed or even cannot be made. Eventually the resident’s lives and reliabilities
cannot be maintained. More proper and quick information service system with disaster
and road environmental states are indispensable.
In order to resolve those problems, we propose IoT based disaster information
platform. In our system, road side wireless nodes and mobile smart nodes with various
sensors and different wireless communication devices organize a largescale information
infrastructure without conventional wired network. The mobile nodes such as cars and
drones collect various sensor data and carry those data as a message ferry while moving
one end to another along the roads.
On other hand, the road side wireless node not only collects and stored sensor data
from its own sensors in its database server but exchanges the sensor data from the
mobile nodes when it passes through the road side wireless node by V2X protocol [3].
Therefore, both sensor data at the road side wireless node and the mobile nodes are
periodically uploaded to cloud system and synchronized. Thus, mobile node perform as
mobile communication means even through the communication infrastructure is
challenged environment or not well developed.
This disaster information platform not only performs various road sensor data
collection and transmission functions, but also performs Internet access network
functions to transmit the various data such as disaster prevention information on
emergency case, road state information, sightseeing information and shopping infor-
mation on ordinal case. Therefore, many applications and services can be realized.
Since the population density in the mountain or rural area is quite low, the commu-
nication infrastructure is not well developed although many residents in those areas
spend daily life. Therefore, it is desired to realize information infrastructure to take
advantage of automobile. Automobile can always move anywhere along roads and
encounter the other automobile. Also by installing the base stations along roads, the
automobile can communicate and exchange to those base stations and eventually
communicate to Internet and cloud computing. Furthermore, even though the road
network infrastructure is well developed, drone can fully or partially perform the
function of automobile. In the future, since the drone improves its flying performance
and expected to launch up from and arrive directly to automobile, more flexible and on
demand information exchange functions can realized.
The Fig. 1 shows a system configuration of our proposed IoT based mobility
information infrastructure which consists of multiple wireless networks and nodes
including fixed road side base station, balloon, mobile vehicle, drone in addition to
satellite and various offices and shelters. Those nodes are mutually connected with
various wireless communication networks including LPWA, 3G, LTE, Wi-Fi, Wi-Max,
satellite in addition to wired networks. In the future, 5G is also connected as wireless
network. The mobile wireless nodes are organized by different typed wireless network
devices and various sensor devices such as camera, semi-electrostatic field sensor,
acceleration sensor, gyro sensor, temperature sensor, humidity sensor, infrared sensor
and sensor server.
A IoT Based Disaster Information Platform 539
Figure 2 shows a fixed wireless node including fixed road side base station, various
offices and shelters which performs sensing environmental conditions such as road
condition, mountain condition and disaster area state using camera, temperature,
humid, rain, wind power, etc. Those fixed nodes also communication function, storage
function of the sensing data and exchange function of the sensor data from the mobile
node and autonomous power supply system which is consisted of window turbine,
solar panel, battery and controller [10, 11, 13].
[12]. If none of link connection is existed, those transmitted sensor data and disaster
information are locally and temporally stored in the storage until another mobile node
approaches, and starts to transmit data by Delay Tolerant Network (DTN) function [4–
6]. Thus, data communication can be attained even though the network infrastructure is
destroyed or not existed in challenged network environment such as mountain areas or
just after large scale disaster areas.
Figure 3 shows mobile wireless node including automobile, drone and balloon
nodes and consists of various environmental sensors, cognitive wireless network with
different frequency bands as the fixed wireless nodes, Software Defined Network
(SDN) function, and Delay Tolerant Network (DTN) function and Storage function.
communication began to connect to neighbor automobile and base station, the best
communication link is selected using SDN function and eventually those accumulated
data are downloaded to the communication node on the ground.
All of the collected data from the sensors are linked to the equivalent the geological tab
such as latitude, longitude, height and time. However, those data are not temporally and
spatially homogeneous because the sensor locations are irregular position nor the
sampling period and duration are not well synchronized. In another words, those data
have both geologically sparse and temporally time series characteristics. Therefore, in
order to process those sensor data to make a decision for the state of environmental
state. In our research, more active measuring method is introduced in which those
nonhomogeneous data can be compensated to the homogeneous data by temporal and
spatial interpolation and extrapolation function and integration and distribution func-
tion by optimizing the sensing granularity in both spatial and time domain.
542 Y. Shibata et al.
6 Disaster Application
The Figs. 5 and 6 show a Tourist and Disaster Information using our proposed system
for normal case and emergent case [7–10]. In normal situation, many Japanese tourists
and foreign tourists from different countries visit to the Points of Interest (POI) such as
the historical places, national parks and mountains and spas where network information
infrastructures are usually not well prepared in snow country in Japan. Even worse, it is
difficult for foreigners to understand the information of Points POI written in Japanese.
However, by introducing the proposed mobility information infrastructure, and using
smart phones and tablet terminals, the tourists can ask and speak by their own voices
and obtain information POI by their own languages, eventually being navigated to the
objective places.
On the other hand, when disaster occurred, the push typed disaster information,
evaluation information are automatically delivered to those tourists by their languages.
Eventually, the tourist can safely evacuate to the proper shelter from the current
location. Through the mobility information infrastructure, the disaster state informa-
tion, resident safety information, required medicine, feeds and materials are also col-
lected and transmitted by mobile nodes between the counter disaster headquarter and
evacuation shelters as shown in Fig. 6.
A IoT Based Disaster Information Platform 543
7 Conclusions
In this paper, a new IoT based disaster information platform is introduced. On the
proposed platform, mobility such as cars and drones can performs information and
communication means to organize an information infrastructure in challenged network
environment where both wired and wireless network are not well installed such as
mountain areas and coast areas in snow countries. In the disaster cases as emergency
situations, drones are launched from cars and can organize temporal wireless mesh
networks in the sky and can cover the wide communication areas. Thus by combing
both the cars and drones, very important information related to resident’s life and
property can quickly and flexibly exchanged using this platform. Currently we are
implementing a prototype of the proposed the disaster information platform by
extending previously developed Never Die Network (NDN) to evaluate the effects and
usefulness of the proposed platform through annual disaster evaluation drill in a snow
country.
References
1. A 2016 Aging Society White Paper. http://www8.cao.go.jp/kourei/whitepaper/w-014/
zenbun/s1_1_1.html
2. A 2016 Declining Birth rate White Paper. http://www8.cao.go.jp/shoushi/shoushika/
whitepaper/measures/english/w-2016/index.html
544 Y. Shibata et al.
3. Ito, K., Hirakawa, G., Shibata, Y.: Experimentation of V2X communication in real
environment for road alert information sharing system. In: IEEE AINA 2015, pp. 711–716
(2015)
4. Otomo, M., Sato, G., Shibata, Y.: In-vehicle cloudlet computing based delay tolerant
network protocol for disaster information system. In: Advances on Broad-Band Wireless
Computing, Communication and Application Applications. Lecture Notes on Data
Engineering and Communications Technologies, vol. 2, pp. 255–266 (2016)
5. Hirakawa, G., Uchida, N., Arai, Y., Shibata, Y.: Application of DTN to the vehicle sensor
platform CoMoSe. In: WAINA 2015, pp. 24–27 (2015)
6. Kitada, S., Sato, G., Shibata, Y.: A DTN based multi-hop network for disaster information
transmission by smart devices. In: Advances on Broad-Band Wireless Computing,
Communication and Application Applications. Lecture Notes on Data Engineering and
Communications Technologies, vol. 2, pp. 601–611 (2016)
7. Shibata, Y., Goto, T., Sato, G., Hashimoto, K.: Disaster information sharing system
considering communication status and elapsed time. In: IWDENS (2017)
8. Shibata, Y., Uchida, N., Shiratori, N.: Analysis and proposal of disaster information network
from experience of the great East Japan earthquake. IEEE Commun. Mag. 52, 44–48 (2014)
9. Japan Police Department: The Great East Japan Disaster. http://www.npa.go.jp/archive/
keibi/biki/index.htm
10. Uchida, N.: Min-max based AHP method for route selection in cognitive wireless network.
In: Proceedings of the 2010 13th International Conference on Network-Based Information
Systems, NBIS 2010, pp. 22–27 (2010)
11. Uchida, N., Takahata, K., Shibata, Y.: Disaster information system from communication
traffic analysis and connectivity (quick report from Japan Earthquake and Tsunami on March
11th, 2011). In: The 14th International Conference on Network-Based Information Systems
(2011)
12. Open Networking Foundation: Intro to OpenFlow. https://www.opennetworking.org/
standards/intro-to-openflow
13. Uchida, N., Sato, G., Shibata, Y., et al.: Selective routing protocol for cognitive wireless
networks based on user’s policy. In: The 12th International Workshop on Multimedia
Network Systems and Applications (MNSA 2010), pp. 112–117 (2010)
Study on Balloon Network Using LoRa Mesh
Communication System
Fig. 1. Ground-to-ground LoRa communication experiments carried out across the bay
2 Methods
For situations like this, where we assume a network connecting centers that are
some distance away from each other, such as office space and evacuation centers and
exchange relatively lightweight data such as safety information, we set the spreading
factor and used channel bandwidth setting to slow But adjust to parameters that allow
longer distance communication.
Evaluate the procedures and construction time of the over-the-air network, human
resource, continuous operation time, and clarify the difference of features of ground
communication and over-the-air communication, advantages and disadvantages.
Measure the effective throughput of the constructed network and the time taken for data
synchronization and evaluate whether it meets the requirements of the application to be
implemented. Since LoRa wireless communication uses broadcast communication, if
collision of wireless signals occurs due to simultaneous transmission, retransmission is
not performed. Also, in consideration of various communication environments, nodes
that perform communication have hidden terminal problems because radio wave reach
ranges are extremely different.
In order to avoid simultaneous transmission, LoRa mesh network technology
implements time-division access control method after time synchronization by GPS. It
is assumed that all the radio stations operated in the system are numbered by ID and the
total number is also grasped. In this premise, based on the ID dynamically allocate time
slots that it can transmit. For example, assuming that the transmission time per wireless
device is 10 s (considering the time required for backoff to avoid collisions with other
systems), when using five wireless stations, these five devices do not have any prob-
lems A time longer than 50 s that can be transmitted is assigned as a transmission
cycle. In this transmission cycle, as shown in Fig. 3, transmission time slots are
assigned based on respective IDs. As a result, occurrence of collision in communication
in the own system is avoided. The amount of data that can be transmitted per unit time
548 G. Sato et al.
of the wireless device and the communicable distance vary depending on factors such
as the spreading factor of LoRa, the used channel bandwidth, the surrounding com-
munication environment, and the like. Also, since the transmission cycle varies
according to the number of radio devices used in the system, slot design of the entire
system is required in advance according to the requirements.
Fig. 3. Relationship between transmission cycle time and transmission time per wireless device
We are prototyping a prototype system (Figs. 4 and 5) that currently implements this
protocol based on NerveNet [1] developed by NICT, and we are carrying out a lab-
oratory test. Basic protocol operation can be confirmed by the expected operation, basic
data collection, system design, and implementation for each module are expected to be
completed in the current fiscal year. We will conduct information transfer experiments
using various data in future and demonstrate communication outdoors.
Study on Balloon Network 549
4 Summary
In this paper, We proposed combining LoRa mesh network system developed by NICT
with overhead communication utilizing balloon and drone. I explained the basic
architecture of the LoRa mesh network system and outlined the balloon network for
applying it. We are also developing a power saving and lightweight wireless module
that can be attached to a balloon as a prototype system. In the future, we will form a
network between wireless modules actually mounted on the balloon and verify how
useful it is compared to ground communication.
Reference
1. Inoue, M., Ohnishi, M., Peng, C., Li, C., Owada, Y.: NerveNet: a regional platform network
for context-aware services with sensors and actuators. IEICE Trans. Commun. E94-B(3),
618–629 (2011)
Performance Evaluations on Adaptive
Array Antenna of Vehicular Delay
Tolerant Networking for Winter Road
Surveillance Systems
1 Introduction
Thus, this paper proposes the optimal AAA (Adaptive Array Antenna) control
methods with the camera image recognition for the DTN (Delay Tolerant Networks)
method in the road surveillance system in the winter [2]. Also, it is especially discussed
about the performance evaluations of the remote voltage control methods for the
antenna elements in order to realize the quick adjustments of the beam-forming for the
target vehicles in this paper. The proposed system consists of the camera recognition
for the target automobiles, the beam-forming antenna that is calculated by the Kalman
Filter Algorithm [3, 4] in order to reduce the radio noises, and the voltage control
device for the antenna elements [5]. Then, this paper newly reports the improved
methods and the experimental results of the rapid voltage control methods for the
Adaptive Array Antenna (AAA) with the asynchronous communication by using the
VISA (Virtual Instrument Software Architecture) standard, and the future research
subjects are discussed.
In the followings, the proposed methods of the AAA with the image recognition for
the winter road surveillance system are introduced, and the prototype system in the
previous study is explained in Sect. 2. Moreover, the proposed rapid voltage controls with
the asynchronous communication by using the VISA are discussed in Sect. 3. Finally, the
experiments for the evaluation of the proposed AAA control methods are discussed in
Sect. 4, and the conclusions and the future researches are discussed in Sect. 5.
2 Proposed Systems
Figure 1 shows the proposed road surveillance system in the winter [2]. In the system,
each vehicle equips the wireless networks such as IEEE802.11p [6], and the camera
system to recognize the target vehicles, and the AAA system to adjust the beam-
forming radio in order to establish the longer transmission of the target vehicles. The
V2V networks are based on the DTN method since the proposed system is assumed to
use in the mountain area where LTE or 3G networks are not provided, and so the
vehicles are possible to receive the road information from the target vehicles through
the V2V communication. It is also assumed that the road information is acquired from
the sensors from each vehicle’s wheel or on each vehicle.
Fig. 1. The proposed systems for the road surveillance system in the winter.
552 N. Uchida et al.
Fig. 2. The target connectivity of the DTN based V2V networks in the paper.
Fig. 3. The prototype system for the winter road surveillance system.
Also, since the rapid controls of the beam-forming, the proposed system introduced
the image recognition for the target vehicles. The previous papers [7, 8] introduced the
AAA controls by the MMSE (Minimum Mean Square Error) or the LMS (Least Mean
Square), but there are the problems of the convergence the complexity for the
Performance Evaluations on Adaptive Array Antenna 553
calculation. Therefore, the AAA controls with the image recognition with the Machine
Learning algorithm is introduced in this paper, and the Kalman Filter algorithm is
introduced the reduction of the radio noises from the rapid movements of the auto-
mobiles on the roads.
Besides, the DTN routing [9, 10] is introduced in the proposed network commu-
nications. The DTN is the stored-and-carried typed protocols for the robust network
conditions such as the interplanetary communications originally, and it is necessary to
consider the mountain areas in the proposed systems. The previous study [11] indicated
that the prototype system of the proposed is shown in Fig. 3. The system mainly
consists of three parts such as the camera system (Logicool Web camera c270) for the
recognition of the target vehicle, the PC (Macbook 12inch) for the calculation of the
distance and the angle for the target vehicle and for the predictions by the Kalman
Filter, and the AAA system that is controlled by the voltage machine (PMX18-5A by
KIKUSUI Corporation). However, there were some problems with the performance of
the directional controls of the AAA system, and this paper will discuss the enhanced
voltage control methods by asynchronous communication.
This paper newly proposes the rapid voltage controls by the asynchronous commu-
nication by using the VISA standard [12]. The VISA is the standard for configuring,
programming, and troubleshooting instrumentation systems comprising GPIB, VXI,
PXI, Serial, Ethernet, and USB interfaces. Therefore, it helps the various programming
environments such as Microsoft Visual Studio or JDK (Java Development Kit) for the
multiple interfaces.
Figure 4 shows the program flow of the proposed asynchronous communication by
using the VISA standard. Although the previous prototype system [11] introduced the
synchronous communication for remoting the voltage machine, it was the problems to
adjust the beam-forming of the proposed AAA system. According to the previous study
indicated, it took more than one second to reach the voltage from zero volts to five
volts. If the vehicle run by 60 km/h, the vehicle will move about 16.7 m per one
second. Therefore, it was necessary to introduce more quick voltage remote system for
the prototype system.
In Fig. 4, first of all, the connection between the PC and the voltage machine is
established by the “OPEN” command of the VISA standard through the LAN. Then,
the remote control mode is activated by the “WriteString” command of the syn-
chronous communication, and the value of the target voltage is sent after receiving the
response from the voltage machine. Then, the current voltage setting is continuously
asking the voltage machine, and the connection is closed when the current voltage is
reached to the target voltage. Although the error functions are the future subjects for
this asynchronous controls, there were certain waiting times in case of the previous
controls, and the improvements of the proposed control methods are discussed in the
following chapters.
554 N. Uchida et al.
Fig. 4. The program flow of the proposed asynchronous communication by using the VISA
standard in order to control the beam-forming of the proposed AAA system.
4 Experiments
The experiments with the prototype system evaluate the effectiveness of the remote
voltage controls by the proposed asynchronous communications. In the experiments,
first of all, the processing time to reach the given voltage is observed by the system.
Table 1 shows the results of the processing time when the given voltage of the VISA
command is set to 1, 3, 5…15 V from 0 V, and the actual voltage output is shown in
the table.
Table 1. Processing time to reach the given voltage by the VISA commands.
Voltage changes Processing time (s) Observed voltage (V)
0 to 1 (V) 0.554 0.99
0 to 3 0.643 2.99
0 to 5 0.700 4.99
0 to 7 0.727 6.99
0 to 9 0.810 9.00
0 to 11 0.831 11.00
0 to 13 0.863 13.00
0 to 15 0.885 14.99
The previous paper [11] present that the processing time was 1.00 s from 0 V to
5 V, 2.70 s from 0 V to 5 V, and 3.69 s from 0 V to 15 V, but the proposed asyn-
chronous controls reached from 0 V to 15 V by 0.88 s in the experiments. According
to the results, it is effective that the proposed methods provide the quick beam-forming
Performance Evaluations on Adaptive Array Antenna 555
controls in the prototype system, and it is suggested that the proposed methods are
useful for the V2V networks even if the vehicles moves so fast.
The reason for the rapid controls by the asynchronous command communication
could be the guaranteed time for waiting for the response from the voltage machine in
the VISA standards for the safety device controls, and so it is necessary to consider the
additional error controls in case of the malfunctions of the voltage machine or LAN
disconnections for the future studies.
Secondly, the processing time to control the given voltages and the angle of the
beam-forming by the prototype AAA system is shown in Fig. 5 and Table 2. Figure 5
indicates the differences between the processing time of the synchronous controls in the
previous study [11] and that of the proposed asynchronous controls.
Fig. 5. Results of the given voltages in the prototype AAA system in the experiments.
In the experiments, the voltage VISA commands were sent from the note PC to the
voltage machine by every 5.00 s, and the processing time of the increasing/decreasing
5.00 V was observed in the prototype system. The figure shows that the proposed
asynchronous controls rapidly reached to the given voltage at every point, and it is
supposed that the proposed methods are effective for the V2V communication. Also,
Table 2 shows the angle of the beam-forming in the AAA system and the processing
time in the experiments. It took only 0.855 s for changing 45° of the beam-forming in
the system, and it is also supposed that the proposed methods are more useful for the
rapid movements of the V2V communications.
556 N. Uchida et al.
Table 2. Processing time to reach the target voltage by the VISA commands.
Setting Voltage (V) Processing time (s) Peak angle of beam-forming (degree)
5 0.700 10.0
10 0.820 30.0
15 0.885 45.0
Acknowledgement. This paper is the extend version of “Remote Voltage Controls by Image
Recognitions for Adaptive Array Antenna of Vehicular Delay Tolerant Networks” in the 13th
International Conference on Advances on P2P, Parallel, Grid, Cloud and Internet Computing
(3PGCIC2018), Oct. 2018. Also, this work was supported by SCOPE (Strategic Information and
Communications R&D Promotion Programme) Grant Number 181502003 by Ministry of
Internal Affairs and Communications in Japan.
References
1. Kato, S., Minobe, N., Tsugawa, S.: Experimental report of inter-vehicle communications
systems: a comparative and consideration of 5.8GHz DSRC, wireless LAN, cellular phone,
IEICE Techn. Rep. ITS 103(242), 23–28 (2012)
2. Uchida, N., Takahata, K., Shibata, Y., Shiratori, N.: Proposal of vehicle-to-vehicle based
delay tolerant networks with adaptive array antenna control systems. In: The 8th
International Workshop on Disaster and Emergency Information Network Systems
(IWDENS2016), pp. 649–654 (2016)
3. Welch, G., Bishop, G.: The Kalman Filter. http://www.cs.unc.edu/*welch/kalman/
Performance Evaluations on Adaptive Array Antenna 557
4. Welch, G., Bishop, G.: An Introduction to the Kalman Filter. University of North Carolina at
Chapel Hill Department of Computer Science (2001)
5. Uchida, N., Ichimaru, R., Ito, K. Ishida, T., Shibata, Y.: Implementation of adaptive array
antenna controls with image recognitions for DTN based vehicle-to-vehicle networks. In:
The 9th International Workshop on Disaster and Emergency Information Network Systems
(IWDENS2017), pp. 633–638 (2017)
6. Aizawa, T., Shigeno, H., Yashiro, T., Matushita, Y.: A vehicle grouping method using W -
CDMA for road to vehicle and vehicle to vehicle communication. IEICE Technical
Report ITS 2000(83(2000-ITS-002)), pp. 67–72 (2000)
7. Karasawa, Y., Sekiguchi, T., Inoue, T.: The software antenna: a new concept of
Kaleidoscopic antenna in multimedia radio and mobile computing era. IEICE Trans.
Commun. E80-B, 1214–1217 (1997)
8. Yokoi, T., Iki, Y., Horikoshi, J., et al.: Software receiver technology and its applications.
IEICE Trans. Commun. E83-B(6), 1200–1209 (2000)
9. Fall, K., Hooke, A., Torgerson, L., Cerf, V., Durst, B., Scott, K.: Delay-tolerant networking:
an approach to interplanetary internet. IEEE Commun. Maga. 41(6), 128–136 (2003)
10. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Proposal of Data Triage Methods for
Disaster Information Network System Based on Delay Tolerant Networking: The 7th
International Conference on Broadband and Wireless Computing, Communication and
Applications (BWCCA 2013), pp. 15–21 (2013)
11. Uchida, N., Kawamura, N., Shibata, Y., Shiratori, N.: Remote Voltage Controls by Image
Recognitions for Adaptive Array Antenna of Vehicular Delay Tolerant Networks: Advances
on P2P, Parallel, Grid, Cloud and Internet Computing. (3PGCIC 2018). Lecture Notes on
Data Engineering and Communications Technologies, vol. 24, pp. 387–394. Springer (2018)
12. National Instruments Coorporation: National Instruments VISA. https://www.ni.com/visa/
The 7th International Workshop on
Collaborative Emerging Systems
(COLLABES-2019)
Design of Modern Logistics Management
System Based on RFID and NB-IoT
Jiayi Pang1 , Leixian Shen1 , Qingyun Zhang1 , He Xu1,2(B) , and Peng Li1,2
1
School of Computer, Nanjing University of Posts and Telecommunications,
Nanjing 210023, China
{b16041505,b16041522,b16041523,xuhe,lipeng}@njupt.edu.cn
2
Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks,
Nanjing 210003, China
1 Introduction
In the current express delivery industry, the barcode is still the main mean of
information storage and identification because of its simplicity and low cost, but
scanning barcodes requires a lot of manpower and material resources. Through-
out the logistics process, the barcode of each cargo needs to be scanned multiple
times. Moreover, because of the short recognition distance, only when the loca-
tion of the barcode is found can it be scanned, and only one can be scanned at
a time.
An express warehouse usually has a large throughput and fast circulation.
The traditional barcode has low efficiency of scanning and cannot support the
high-speed operation of the logistics industry. The RFID technology has the
advantages of non-contact, large data volume and long recognition distance, so
this paper proposes a logistics management system which replaces barcodes with
RFID tags.
2 Related Work
RFID can be applied in many aspects of logistics and will make great contribu-
tions to the sorting of cargoes and management of warehouses, which can greatly
improve the degree of automation and improve efficiency.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 561–569, 2019.
https://doi.org/10.1007/978-3-030-15035-8_54
562 J. Pang et al.
Cargo sorting is usually carried out on conveyor belts and requires the posi-
tioning of mobile RFID tags. There are many ways to identify and track RFID-
attached cargoes on conveyor belts. In [5], a new synthetic array phase-based
technique for localization and tracking of items equipped with UHF-RFID tags
are investigated. In [12], authors combining received signal strength indication
(RSSI) and Doppler frequency shift measurements to realize sorting for the con-
veyor belt scene of dense tags. It can be seen that the use of RFID RSSI and
other attributes to achieve cargo sorting is feasible and effective. In practical
applications, RFID-based airport bag sorting has been implemented.
When using RFID to manage warehouses, the core goal is to use RFID to
locate cargoes, thereby achieving forklift route optimization, handling warehouse
operation orders smartly, cargo inventory, etc. Cargo positioning is a kind of
indoor positioning, often using RSSI-based positioning algorithm [8–10], or using
phase-based algorithm [3,7,13], and in recent years there are also methods of
using machine learning [6,11]. NB-IoT has a lot of room for development as a
new technology. Due to its small bandwidth, long battery life and wide coverage,
it has great advantages in smart logistics [1,2].
3 Design of System
This section discusses the design of the whole system, including the system
architecture, data transmission and storage scheme and four logistics-related
schemes based on RFID.
Fig. 1. System architecture diagram. The layers have different components and play
different roles and they are interconnected to form the entire system.
on the corresponding shelf according to the received short message, and scans
the cargoes on the self-service exit. Export’s camera takes a photo of a valid
document or face while scanning. Finally, when the staff needs to inventory the
cargoes, the RFID cargo positioning system will compare the positioning results
with the database information, and promptly find out the cargo that fails to be
scanned but the database shows that it has not been signed.
This scheme builds a narrow-band IoT network based on RFID reader, Rasp-
berry pie processor, NB-IoT module, operator platform, server, database. This
solution builds a logistics data transmission and storage solution, mainly based
on NB-IoT network, supplemented by WLAN and Ethernet networks.
the sorting result on the electronic screen, or calls the audio file to broadcast the
sorting result to the sorter, and finally sorts manually.
The actual cargo positioning scenario is shown in Fig. 7. After obtaining the
positioning information of a cargo tag, the information of each cargo tag read
out is compared with the reference tag information, and the position on the shelf
can be judged according to the model. Finally, Update the location information
of cargoes in the database.
After the staff put cargoes on the shelf, the RFID antenna collects the tag
information on cargoes and transmits them to the cargo positioning system. The
cargo positioning system uses the algorithm to obtain the shelf where cargoes
are located, and automatically sends a pickup notification message according to
the recipient information stored in the tag information, and the consignees can
self-service receive cargoes. At the same time of locating cargoes, the system
performs the express delivery at a certain frequency. For the express delivery
that has not been scanned and has not been signed, the staff is immediately
reminded to check whether cargoes are lost.
5 Conclusions
First of all, the modern logistics management system proposed in this paper
uses the express mail with RFID tags, which greatly improves the automation
of the whole process of logistics. Secondly, the system uses NB-IoT technol-
ogy as the end communication method, which can reduce the communication
cost, and is also the development and expansion of NB-IoT technology. Finally,
replacing barcodes with RFID tags can lay the foundation for fully automated
warehousing and unattended distribution. Under the background of the Internet
of Things, the logistics management system proposed in this paper is in line with
the inevitable trend of smart logistics development, and truly realizes the goal
of serving the society with science and technology.
Design of Modern Logistics Management System 569
References
1. Adhikary, A., Lin, X., Wang, Y.P.E.: Performance evaluation of NB-IoT coverage.
In: 2016 IEEE 84th Vehicular Technology Conference (VTC-Fall), pp. 1–5. IEEE
(2016)
2. Chen, J., Hu, K., Wang, Q., Sun, Y., Shi, Z., He, S.: Narrowband internet of things:
implementations and applications. IEEE Internet Things J. 4(6), 2309–2314 (2017)
3. Hekimian-Williams, C., Grant, B., Liu, X., Zhang, Z., Kumar, P.: Accurate localiza-
tion of RFID tags using phase difference. In: 2010 IEEE International Conference
on RFID, pp. 89–96. IEEE (2010)
4. Huai, Y., Shao, Q., Lu, Z.: Application research of UM220 Beidou/GPS hybrid
position module. Foreign Electron. Meas. Technol. 33(03), 76–79 (2014)
5. Nepa, P., Lombardini, F., Buffi, A.: Location and tracking of items moving on a
conveyor belt and equipped with UHF-RFID tags. In: 2012 IEEE Antennas and
Propagation Society International Symposium (APSURSI), pp. 1–2. IEEE (2012)
6. Ng, W.W., Lin, L., Chan, P.P., Yeung, D.S.: 3D goods allocation in warehouse
with L-GEM based 3-D RFID positioning. In: 2011 International Conference on
Machine Learning and Cybernetics (ICMLC), vol. 1, pp. 324–329. IEEE (2011)
7. Shangguan, L., Yang, Z., Liu, A.X., Zhou, Z., Liu, Y.: Relative localization of RFID
tags using spatial-temporal phase profiling. In: NSDI, pp. 251–263 (2015)
8. Siachalou, S., Bletsas, A., Sahalos, J., Dimitriou, A.G.: RSSI-based maximum like-
lihood localization of passive RFID tags using a mobile cart. In: 2016 IEEE Wireless
Power Transfer Conference (WPTC), pp. 1–4. IEEE (2016)
9. Subedi, S., Pauls, E., Zhang, Y.D.: Accurate localization and tracking of a passive
RFID reader based on RSSI measurements. IEEE J. Radio Freq. Identif. 1(2),
144–154 (2017)
10. Wang, G., Qian, C., Shangguan, L., Ding, H., Han, J., Yang, N., Xi, W., Zhao,
J.: HMRL: relative localization of RFID tags with static devices. In: 2017 14th
Annual IEEE International Conference on Sensing, Communication, and Network-
ing (SECON), pp. 1–9. IEEE (2017)
11. Xu, H., Ding, Y., Li, P., Wang, R.: Indoor localization using the reference tags and
phase of passive UHF-RFID tags. Int. J. Bus. Data Commun. Network. (IJBDCN)
13(2), 69–82 (2017)
12. Xu, J., Yang, L., Liu, Q., Hu, J., Song, T.: A sorting algorithm for RFID tags mov-
ing on a conveyor belt. In: 2018 IEEE International Conference on RFID (RFID),
pp. 1–7. IEEE (2018)
13. Yang, L., Chen, Y., Li, X.Y., Xiao, C., Li, M., Liu, Y.: Tagoram: real-time tracking
of mobile RFID tags to high precision using cots devices. In: Proceedings of the
20th Annual International Conference on Mobile Computing and Networking, pp.
237–248. ACM (2014)
Evaluation of TBC and OBC Precedent
Relations Among Messages in P2P Type
of Topic-Based Publish/Subscribe System
1 Introduction
2 System Model
In P2PPS (peer-to-peer publish/subscribe) systems [1,11–13,17], peer processes
(peers) p1 , ..., pn (n ≥ 1) are cooperating with one another by exchanging mes-
sages in underlying networks. Here, there is no centralized coordinator [3] to
do the communication among peers. Each peer pi can publish messages while
receiving messages from other peers. A message m is only received by a target
peer which is interested in the contents of the message m.
In this paper, peers exchange objects with one another by publishing and
receiving messages. Contents [4,7] of a message m are considered to be a collec-
tion of objects carried by the message m. Let O be a set of objects in a system.
An object is a unit of data resource. Let m.O (⊆ O) be a set of objects carried
by a message m. A peer which is interested in the contents m.O of a message
m is a target peer of the message m. A peer pi supports a memory pi .M to
store objects. A peer pi creates and manipulates objects in the memory pi .M .
On receipt of a message m, a peer pi stores objects m.O carried by the message
m in the memory pi .M . In this paper, a pair of object collections O1 and O2 are
related if O1 ∩ O2 = φ, i.e. O1 and O2 include a common subobject. A transac-
tion is a unit of work of each peer pi . A transaction is a sequence of publication
and receiving events of messages. A peer pi is modeled to be a sequence of trans-
actions Ti1 , ..., Tili . This means, objects are not persistent. Messages published
and received in a pair of different transactions Tik and Tih are not related. That
is, each time a transaction starts on a peer pi , the memory pi .M is flushed.
We assume the underlying network supports a pair of peers pi and pj with
the following communication service:
1. Each peer pi receives only and all messages in whose contents the peer pi is
interested. Here, the peer pi is a target peer of the message m.
2. Messages published by each peer can be received by every target peer in the
publishing order without message loss and duplication.
3. Every message published by a peer arrived at every peer.
If a peer pi publishes a message m1 before another message m2 , every com-
mon target peer pj receives the messages m1 then m2 in the publishing order.
However, some pair of messages m1 and m2 published by different peers may not
be received by every common destination peer in the same order. For example,
a peer pi publishes a message mi1 then mi2 and another peer pj publishes a
message mj1 then mj2 . Suppose a pair of peers pk and pl are common target
peers of the messages mi1 , mi2 , mj1 , and mj2 . Here, the target peer pk receives
the messages mi1 , mj1 , mj2 , and mi2 in the sequence. On the other hand, the
other peer pl receives the messages in a sequence mj1 , mi1 , mj2 , and mi2 . Here,
each peer receives a pair of messages mi1 and mi2 from the peer pi and the
messages mj1 and mj2 from other the peer pj in the publishing order. However,
the peer pk receives mi1 before mj1 but the peer pl receives mj1 before mi1 . If
the message mi1 causally precedes the message mj1 (mi1 → mj1 ), the peer pl
has to deliver mi1 before mj1 as the peer pk . Each peer has to causally deliver
only and every pair of related messages.
A Causally Precedent Relation 573
Fig. 1. OBC-precedency.
and pk receives both the messages mi and mj . Here, the peer ph first receives
the message mi and then the message mj . On the other hand, another peer pk
first receives the message mj and then the message mi in an order different from
the peer ph . Thus, messages published by different peers may be received by
different common target peers in different orders depending on network delay
time because there is no centralized coordinator. If the message mi OBC-precedes
the message mj (mi ⇒ mj ), each common target peer pk is required to receive
the message mi before the message mj . However, it is not easy for each peer
to hold every object received and manipulate objects to check if the objects are
related. In this paper, we newly define a topic-based-causally (T BC) precedent
relation.
[Definition] Let m1 and m2 be a pair of messages and B be a subset of topics
(B ⊆ m1 .P T ∩ m2 .P T ).
1. m1 primarily T BC-precedes m2 with respect to B (m1
→B m2 ) iff m1 is
received and then m2 is published before m2 .
2. m1 T BC-precedes m2 with respect to B (m1 B m2 ) iff one of the following
conditions holds:
• m1
→B m2 .
• m1
→B m3 and m3 B m2 for some message m3 .
3. m1 is TBC-concurrent with m2 with respect to B (m1 |||B m2 ) iff nether m1
B m2 nor m2 B m1 .
Suppose m1
→B1 m2
→B2 ...
→Bk−1 mk . Here, it is noted m1 B mk may
not hold. m1 B mk means that the topic sets B1 , ..., Bk−1 include common
topics B, i.e. B ⊆ B1 ∩ ... ∩ Bk−1 .
A message m1 TBC-precedes a message m2 (m1 m2 ) iff m1 B m2
for some subset B (⊆ T ) of topics. A pair of messages m1 and m2 are TBC-
concurrent (m1 ||| m2 ) iff m1 |||B m2 for some subset B (⊆ T ).
576 T. Saito et al.
Suppose there are four peers pi , pj , ph , and pk as shown in Fig. 3. The peer pi
first publishes a message mi . The peer pj publishes a message mj after receiving
the message mi . Lastly, the peer pk publishes a message mk after receiving the
message mj . Suppose mi .P T = {a, b}, mj .P T = {a, c}, and mk .P T = {a, d}.
Here, the message mi primarily TBC-precedes the message mj (mi
→{a} mj )
since mi → mj and {a} ⊆ mi .P T ∩ mj .P T . Hence, the message mi TBC-
precedes a pair of the messages mj and mk with respect to a topic a, i.e. mi
{a} mj and mi {a} mk . It is noted, each of the messages mi , mj , and mk
includes the publication topic a and mi
→{a} mj and mj
→{a} mk .
m carries the linear time m.L when a peer sends the message m. On receipt of
a message m, the liner time L of a peer pi is max (L, m.L). In order to publish
a message m, a peer pi increments the variable L by one, i.e. L = L + 1, and
m.L = L. Then, the peer pi publishes the message m. Suppose a peer pi receives
a pair of messages m1 and m2 . If m1 .L < m2 .L, the message m1 is delivered
before the message m2 in the target peer pi . Otherwise, the messages m1 and
m2 are delivered in the receipt order.
Suppose a peer pi publishes a message m with publication m.P T . A message
m carries a topic vector m.T V and linear time m.L. The message m is published
by the peer pi as follows:
[Publication at a peer pi ]
5 Evaluation
We evaluate the TBC protocol in terms of number of pairs of messages unnec-
essarily ordered with respect to topics. A system includes peers p1 , ..., pn
(n ≥ 1). Let T be a set of topics t1 , ... tl (l ≥ 1). Let O be a set of objects
o1 , ... ol (l ≥ 1).
Each peer pi randomly publishes messages. That is, the publishing time
m.P BT of a message m published by a peer pi is randomly decided from time 0
to time maxT [time unit (tu)]. In the evaluation, value of maxT is 1000. At each
time unit, a peer pi is randomly decided in all the peers. Then, the peer pi cre-
ates one message m and publishes the message m. The receiving time m.RV Tj
of a message m of each peer pj is decided as m.RV Tj = m.P BT + δij . Here,
the delay time δij between a pair of peers pi and pj is randomly decided from
1 to 10 [tu]. Each message m carries the topic vector m.T V in addition to the
linear time m.L for evaluation.
In the evaluation, publication and subscription topics of each peer pi are
randomly taken from the topic set T . Each peer pi creates only one object o
where the publication o.T is a subset of subscription topics of the peer pi . Then,
the peer pi creates an object o in the set O with probability 0.01. Publication
topics of a message m are topics of objects m.O. In the evaluation, each peer
pi manipulates the vector time variable V T [9] and each message m carries the
vector time m.V T to check the causally precedent relation → among messages.
Each peer pi includes the vector time V T in a message m as m.V T . In the
vector time, “mi .V T < mj .V T ” iff a message mi causally precedes a message
mj (mi → mj ). Even if mi .LT < mj .LT holds, mi → mj may not hold. Here,
if a message mj is received before another message mi , a pair of the messages
mi and mj are unnecessarily ordered.
Suppose a peer pi receives a message m1 before a message m2 but m2 ≺
m1 . In the TBC protocol, the messages m1 and m2 are exchanged in the receipt
queue. Here, if m2 .V T < m1 .V T does not hold, the messages m1 and m2 are
unnecessarily ordered. In the simulation, we obtain the number of pairs of mes-
sages unnecessarily ordered.
Figure 5 shows the number of pairs of messages unnecessarily ordered. Here,
there are 10 peers (n = 10) and 20 topics (l = 20). Subscription of each peer is
randomly selected 1 to 5 topics from all topics (mstn = 5). The number of pairs
of messages unnecessarily ordered linearly increases for the number of messages
published by peers.
580 T. Saito et al.
Fig. 5. Evaluation.
6 Concluding Remarks
References
1. Google alert. http://www.google.com/alerts
2. Google cloud pub/sub. https://cloud.google.com/pubsub/docs/overview?hl=ja
3. Arridha, R., Sukaridoho, S., Pramadihanto, D., Funabiki, N.: Classfiation IoT-big
data analytic for smart environment monitoring and analytic in real-time system.
Int. J. Space-Based Situated Comput. (IJSSC) 7(2), 82–93 (2017)
4. Bouramoul, A.: Contextualisation of information retrieval process and document
ranking task in web search tools. Int. J. Space-Based Situated Comput. (IJSSC)
6(2), 74–89 (2016)
5. Comer, D.E.: Internetworking with TCP/IP (VOLUME I). Prentice Hall, Upper
Saddle River (1991)
A Causally Precedent Relation 581
6. Eugster, P., Felber, P., Guerraoui, R., Kermarrec, A.: The many faces of pub-
lish/subscribe. ACM Comput. Surv. 35(2), 114–131 (2003)
7. Jacobson, V., Smetters, D., Thornton, J., Plass, M., Briggs, N., Braynard, R.:
Networking named content. In: Proceedings of the 5th International Conference
on Emerging Networking Experiments and Technologies (CoNEXT 2009), pp. 1–
12 (2009)
8. Lamport, L.: Time, clocks, and the ordering of event in a distributed systems.
Commun. ACM 21(7), 558–565 (1978)
9. Mattern, F.: Virtual time and global states of distributed systems. In: Parallel and
Distributed Algorithms, pp. 215–226 (1988)
10. Nakamura, S., Enokido, T., Takizawa, M., Ogiela, L.: An information flow control
models in a topic-based publish/subscribe systems. J. High Speed Netw. 24(3),
243–257 (2018)
11. Nakayama, H., Duolikun, D., Enokido, T., Takizawa, M.: Selective delivery of event
messages in peer-to-peer topic-based publish/subscribe systems. In: Proceedings of
the 18th International Conference on Network-Based Information Systems (NBiS-
2015), pp. 379–386 (2015)
12. Nakayama, H., Duolikun, D., Enokido, T., Takizawa, M.: Reduction of unnecessar-
ily ordered event messages in peer-to-peer model of topic-based publish/subscribe
systems. In: Proceedings of IEEE the 30th International Conference on Advanced
Information Networking and Applications (AINA-2016), pp. 1160–1167 (2016)
13. Nakayama, H., Ogawa, E., Nakamura, S., Enokido, T., Takizawa, M.: Topic-based
selective delivery of event messages in peer-to-peer model of publish/subscribe sys-
tems in heterogeneous networks. In: Proceedings of the 18th International Confer-
ence on Network-Based Information Systems (WAINA-2017), pp. 1162–1168 (2017)
14. Raynal, M.: Distributed Algorithms for Message-Passing Systems. Springer,
Heidelberg (2013)
15. Saito, T., Nakamura, S., Enokido, D.D.T., Takizawa, M.: Object-based selective
delivery of event messages in topic-based publish/subscribe systems. In: Proceed-
ings of the 13th International Conference on Broadband and Wireless Computing,
Communication and Applications (BWCCA-2018), pp. 444–455 (2018)
16. Saito, T., Nakamura, S., Enokido, T., Takizawa, M.: A causally precedent rela-
tion among messages in topic-based publish/subscribe systems. In: Proceedings of
the 21st International Conference on Network-Based Information Systems (NBiS-
2018), pp. 543–553 (2018)
17. Setty, V., van Steen, M., Vintenberg, R., Voulgais, S.: PolderCast: fast, robust,
and scalable architecture for P2P topic-based Pub/Sub. In: Proceedings of
ACM/IFIP/USENIX 13th International Conference on Middleware (Middleware-
2012), pp. 271–291 (2012)
18. Tarkoma, S.: Publish/Subscribe System: Design and Principles, 1st edn. John
Wiley and Sons, Ltd. (2012)
19. Tarkoma, S., Rin, M., Visala, K.: The publish/subscribe internet routing paradigm
(PSIRP): designing the future internet architecture. In: Future Internet Assembly,
pp. 102–111 (2009)
Data Analytics for Electricity Load
and Price Forecasting in the Smart Grid
Abstract. The present strategies for the prediction of price and load
may be difficult to deal with huge amount of load and price data. To
resolve the problem, three modules are incorporated within the model.
Firstly, the fusion of Decision Tree (DT) and Random Forest (RF) are
used for feature selection and to remove the redundancy among fea-
ture. Secondly, Recursive Feature Elimination (RFE) is taken for feature
extraction purpose that extracts the principle components and also used
for dimensionality reduction. Finally, to forecast load and price, Support
Vector Machine (SVM) and Logistic Regression (LR) as a classifiers are
used through which we achieve good accuracy results in load and price
prediction.
1 Introduction
2 Related Work
In the operations of modern electricity power system there is the choice of load
shifting as well as we can reduce the electricity energy cost. Dynamic electricity
pricing [7] and [8] helps the user to switch the load.
There are two main ways for the forecasting of load and price i.e. time-
series and machine learning model. For electricity load forecasting [9] developed
a model for the prediction of load and temperature of day ahead electricity
market with the analysis of singular spectrum and neural network structure.
The users are well pleased by knowing that the prices of electricity transcend
the thresholds, on the basis of which they can make the decision to whether turn
off or turn on the load. In the current situation, it is necessary to implement
price data analytic for the classification of electrical price and well as load. There
are various methods to apply classification for load and price forecasting i.e. by
584 S. Aimal et al.
3 System Model
To enquire the problem of electricity load and price forecast. Our aim is to
predict the electricity load and price by using the data from grid and to achieve
good accuracy. To get over this issue, LR and SVM are taken as classifiers that
is used to predict the load as well as price. LR is a type of regression that is use
to predict the occurrence probability of an event, which is done when the data
is fitted to a logistic function. It may use several predictor variables that can be
either categorical or numerical. Historical data is used as a training-set on which
testing is done (Fig. 1). In the proposed approach, we first took the real data-set
Electricity Load and Price Forecasting in Smart Grid 585
from the electricity market (NYISO) i.e. is for three years from january 2011
till 2013. The required data is divided into month-wise and the data with the
same months are utilize for training and testing purpose. For example, for the
training purpose first three weeks of january from 2011 till 2013 are selected and
for the testing purpose last week of january 2011 till 2013 is selected. It consists
of following steps:
• First of all, for the suitable structure format it is necessary to format the
data.
• Incomplete and redundant values and variables are removed.
586 S. Aimal et al.
• The data sampling must be done to reduce consumption time taking by run-
ning algorithm.
The prepared data is used for training, validation, and testing. In this model,
feature extraction, selection, and classification is done. A fusion of DT, RF and
RFE is applied to the data-set for selecting and extracting the principle compo-
nents. RFE is used to eliminate the further redundancy in the selected features.
So, we can say that modules in this model consists of three parts: feature engi-
neering i.e. (selection, extraction) and classification.
The primary goal of the model is to calculate the efficiency and accuracy
while predicting the electricity load and price. For this purpose, we first took
real data from grid, selected features are extracted out and then parameters are
tuned carefully.
The following metrics are the most essential metrics for accessing the per-
formance of the model:
• The model that is applied for the purpose of predicting load and price should
work efficiently and run rapidly.
• Accuracy of classification is influence directly by the performance of feature
engineering.
• Classification’s accuracy is highly dependent on the goal of model.
The main issue that we are facing in the prediction of price and load is
accuracy due to the factors influencing them. However, it makes the training
tough. To enhance prediction accuracy, we embedded feature engineering process
(selection, extraction) in our model.
After feature selection, feature extraction is done which eliminate the redun-
dancy among features using RFE.
3.3 Classifier
• y.pred = classifier.predict(X.test)
Making the confusion matrix
• Simulation Setup.
• Simulation Results.
• Evaluation Performance.
5 Conclusions
In this paper, we’ve enquired the problem of prediction for electricity load and
price in smart grid. To solve the problem, forecasting model is used that is easy to
implement on distributed and parallelized systems. The proposed model is based
on feature engineering and classification. A fusion of two techniques is applied
for feature selection. To further remove the redundancy among features RFE
is used to extract the principle features. Now we have the new features which
enhances the classifiers and achieves the accuracy. LR as a classifier outperforms
best at the forecast accuracy of electric load as well as price. The accuracy that
is achieved by LR is 90%. However, if we see SVM it is not performing the best
as compared to LR. The performance is evaluated by comparing the results with
four means. i.e. MAPE, RMSE, MAE and MSE. For now, we have predicted
load and price for one month; in future we will attempt long term prediction.
i.e. for nine months as well as 12 months.
References
1. Zhang, D., Li, S., Sun, M., O’Neill, Z.: An optimal and learning-based demand
response and home energy management system. IEEE Trans. Smart Grid 7(4),
1790–1801 (2016)
2. Javaid, N., Hafeez, G., Iqbal, S., Alrajeh, N., Alabed, M.S., Guizani, M.: Energy
efficient integration of renewable energy sources in the smart grid for demand side
management. IEEE Access 6, 77077 (2018)
3. Jindal, A., Singh, M., Kumar, N.: Consumption-aware data analytical demand
response scheme for peak load reduction in smart grid. IEEE Trans. Ind. Electron.
(2018)
Electricity Load and Price Forecasting in Smart Grid 591
4. Wang, K., Xu, C., Guo, S.: Big data analytics for price forecasting in smart grids.
In: 2016 IEEE Global Communications Conference (GLOBECOM), pp. 1–6. IEEE
(2016)
5. Wang, K., Xu, C., Zhang, Y., Guo, S., Zomaya, A.: Robust big data analytics for
electricity price forecasting in the smart grid. IEEE Trans. Big Data (2017)
6. Moon, J., Kim, K.H., Kim, Y., Hwang, E.: A short-term electric load forecasting
scheme using 2-stage predictive analytics. In: 2018 IEEE International Conference
on Big Data and Smart Computing (BigComp), pp. 219–226. IEEE (2018)
7. Keles, D., Scelle, J., Paraschiv, F., Fichtner, W.: Extended forecast methods for
day-ahead electricity spot prices applying artificial neural networks. Appl. Energy
162, 218–230 (2016)
8. Chen, P.C., Kezunovic, M.: Fuzzy logic approach to predictive risk analysis in
distribution outage management. IEEE Trans. Smart Grid 7(6), 2827–2836 (2016)
9. Mujeeb, S., Javaid, N., Akbar, M., Khalid, R., Nazeer, O., Khan, M.: Big data
analytics for price and load forecasting in smart grids. In: International Conference
on Broadband and Wireless Computing, Communication and Applications, pp. 77–
87. Springer, Cham (2018)
10. Javaid, N., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
I.A.: A hybrid genetic wind driven heuristic optimization algorithm for demand
side management in smart grid. Energies 10(3), 319 (2017)
11. Mahmood, D., Javaid, N., Ahmed, I., Alrajeh, N., Niaz, I.A., Khan, Z.A.: Multi-
agent-based sharing power economy for a smart community. Int. J. Energy Res.
41, 2074 (2017)
12. Zhao, Z., Lee, W.C., Shin, Y., Song, K.B.: An optimal power scheduling method for
demand response in home energy management system. IEEE Trans. Smart Grid
4(3), 1391–1400 (2013)
13. Logenthiran, T., Srinivasan, D., Shun, T.Z.: Demand side management in smart
grid using heuristic optimization. IEEE Trans. Smart Grid 3(3), 1244–1252 (2012)
14. Ahmad, A., Javaid, N., Alrajeh, N., Khan, Z.A., Qasim, U., Khan, A.: A modified
feature selection and artificial neural network-based day-ahead load forecasting
model for a smart grid. Appl. Sci. 5(4), 1756–1772 (2015)
15. Ahmed, M.S., Mohamed, A., Khatib, T., Shareef, H., Homod, R.Z., Ali, J.A.: Real
time optimal schedule controller for home energy management system using new
binary backtracking search algorithm. Energy Build. 138, 215–227 (2017)
16. Ahmad, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and
fast converging short-term load forecasting model for industrial applications in a
smart grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
17. Wu, M., Wang, Y.: A feature selection algorithm of music genre classification based
on ReliefF and SFS. In: IEEE International Conference on Computer and Informa-
tion Science (ICIS), pp. 539–544 (2009). Processing and Communications Appli-
cations, 2009, pp. 61–64
18. Wu, M., Wang, Y.: A feature selection algorithm of music genre classification based
on ReliefF and SFS. In: IEEE International Conference on Computer and Infor-
mation Science (ICIS), pp. 539–544 (2015)
19. Fleury, A., Vacher, M., Noury, N.: SVM-based multimodal classification of activ-
ities of daily living in health smart homes: sensors, algorithms, and first experi-
mental results. IEEE Trans. Inf. Technol. Biomed. 14(2), 274–283 (2010)
20. Huang, D., Zareipour, H., Rosehart, W.D., Amjady, N.: Data mining for electricity
price classification and the application to demand-side management. IEEE Trans.
Smart Grid 3(2), 808–817 (2012)
An Efficient CNN and KNN Data
Analytics for Electricity Load Forecasting
in the Smart Grid
1 Introduction
Nowadays data analytic plays significant role in every field. Smart grids intro-
duced an advance informational infrastructure that allows optimization in energy
production, transmission, distribution and storage. Smart grid collects data from
different resources and stores it using data analytic techniques. In Smart grid,
bi-directional communication generates data having huge volume, variety and
velocity. Advance data analytic techniques are needed to figure out the required
information from the available raw data. Machine learning techniques are one
of the reliable methods for predicting load and price. Supervised learning algo-
rithms are the part of machine learning which are used for predictions. Cate-
gories of machine learning algorithms are classification and regression. KNN and
CNN are the part of supervised learning algorithm. Deep learning is a learning
method for multiple levels of abstraction and representation which contains an
input layer, hidden layer, and an output layer. Deep learning comes under an
umbrella of machine learning which is based on learning levels of representation,
corresponding to a hierarchy of features, factors or concepts, where the concept
of higher-level is derived from lower-level and vice versa.
The recent advancements in sensor based technologies hold the promise of
novelty in energy consumption models that can better characterize the underly-
ing patterns. Demand-side Management (DSM) is the main and essential com-
ponent towards energy efficiency and management. DSM in smart homes and
residential buildings is a key component towards sustainability and efficiency in
urban environment.
Smart grid is an electrical grid which consists of range of energy and opera-
tional measures. The main objective of grid is to scale back electricity peak load
and to attain the balance between power supply and demand. Customers are
encouraged to partake within the operations of smart grid, the cost of electric
load will be reduced by the preservation of electricity and by shifting the load
towards off-peak hours.
1.2 Contribution
In this paper, We have proposed new classification techniques (i.e. ECNN and
EKNN). The terms forecasting and prediction are used alternatively.
Load is predicted using the new classifiers which beats the traditional tech-
niques in terms of accuracy. Prediction is done on NYISO Data-set. We incorpo-
rated three modules in the proposed methodology. The proposed model consists
of feature engineering and classification. Feature engineering is a two-step pro-
cess (feature selection and feature extraction); for the purpose of feature selec-
tion Mutual Information (MI) is used which reduces the redundancy among fea-
tures and for feature extraction Recursive Feature Elimination (RFE) is used to
extract the principle features from the selected features and reduces the dimen-
sionality of features. Finally, after training the data-set and the removal of the
duplicate features load prediction is done by ECNN and EKNN. The ECNN and
EKNN outperforms better then traditional Convolutional Neural Network and
K-nearest Neighbour.
594 S. Aimal et al.
2 Related Work
In this section, we have discussed different proposed schemes and techniques for
price and load forecasting [1]. The accuracy in forecasting helps the energy sup-
pliers to optimize their strategies for gaining profit. Likewise, to gain a successful
trade in electricity market different techniques have been proposed by different
authors. Short-term Price Forecasting (STPF) and Short-term Load Forecasting
(STLF) play principle role in the management of energy and resource trading
in the grids [2]. Similarly, Auto-regressive Integrated Moving Average (ARIMA)
performs good in stable electricity market.
In the electricity market, load and price forecasting are essential for optimal
operation planning; most of the load and price forecasting methodologies suffer
from lack of efficient feature selection techniques. To overcome this issue, the
authors in [3] proposed a new hybrid filter-wrapper approach that selects the
most relevant features according to the costumer input in a coordinated manner.
In [4], to construct a forecasting model authors presented a hybrid algorithm
which concatenate Empirical Mode Decomposition (EMD), Similar Days (SD)
Selection, and Long Short-term Memory (LSTM) neural network. Moreover,
to evaluate the accuracy of model between the forecasting and historical data
extreme gradient boosting-based weighted k-means algorithm is adopted.
The authors also deal with two types of forecasting online and day-ahead
hourly forecasting [5]. Stacked Denoising Auto-encoder (SDA) models are com-
pared with data-driven approaches in online forecasting whereas in day-ahead
forecasting, the accuracy and effectiveness is further validated. Mathematical
and statistical models like ARIMA, SARIMA etc., belongs to the classical type
of classifiers whereas artificially intelligent models are CNN, LSTM [6] (Table 1).
The authors in [7], proposed an Improved Elman Neural Network (IENN)
based forecast engine to forecast the load. The weights for this prediction model
carried out with an intelligent algorithm to find better forecasting results.
The efficiency of this model is brought-out to the real world engineering in
the comparison with other forecasting models [8]. In [9], the authors focused
on coordinated management of renewable and traditional energy in which the
Electricity Load Forecasting 595
3 System Model
In the proposed approach, we enquire the problem of electricity load prediction.
In our prediction model, we tunned the hyper-parameters of KNN. In addition
to this, we also enhanced CNN in which we used two MAX-pooling and three
convolutional layers. Dense and flatten are used as a fully connected layers.
Historical load data that is taken from grid (NYISO) is used as a training data-
set for EKNN and ECNN; we used 80% of data for training purpose and rest of
20% is used for testing (Fig. 1).
Electricity Load Forecasting 597
1. Feature selection
2. Feature extraction
3. Classification
• The data sampling must be done for the reduction of consumption time.
• Data formatting must be done carefully.
• Redundant data must be removed.
• Proposed model that is applied for the purpose of predicting load must work
efficiently.
• Accuracy of classification can be influenced directly by the performance of
feature engineering.
• Forecasted accuracy is highly dependent on the goal of model.
598 S. Aimal et al.
3.1.3 Classifier
In Classification, we discover a new set of observations on the basis of a trained
set of data. For load prediction, we used ECNN and EKNN as a classifiers.
Electricity Load Forecasting 599
• Setup.
• Results.
• Evaluation Performance.
TM
1
M SE = (Av − Fv )2 (3)
T tm=1
N
n=1 |(Fv − Av )|
M AE = (4)
N
602 S. Aimal et al.
5 Conclusions
In this paper, the problem of electricity load forecast is taken into consideration.
The data is obtained from grid (NYISO data-set) for one year (i.e. january 2016
to 2017). We took required data from the grid which is hourly-base data that
is further utilize for the purpose of training and testing. The process of feature
engineering is done in which the important features are selected using MI. After
the selection of important features, RFE is used for extracting five principle
features.
The trained data-set is further used for load prediction that is done by our
proposed classifiers (ECNN accuracy is 92% and for EKNN accuracy is 93%).
The predicted electricity load outperforms better then other classifiers in terms
of accuracy. The forecast performance is evaluated by comparing the results with
MAPE, RMSE, MAE and MSE. i.e. 10.8,7.5, 7.15, and 10.4 respectively.
References
1. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
2. Liu, B., Nowotarski, J., Hong, T., Weron, R.: Probabilistic load forecasting via
quantile regression averaging on sister forecasts. IEEE Trans. Smart Grid 8(2),
730–737 (2017)
3. Wang, K., Xu, C., Zhang, Y., Guo, S., Zomaya, A.: Robust big data analytics for
electricity price forecasting in the smart grid. IEEE Trans. Big Data (2017)
4. Zheng, H., Yuan, J., Chen, L.: Short-term load forecasting using EMD-LSTM neu-
ral networks with a Xgboost algorithm for feature importance evaluation. Energies
10(8), 1168 (2017)
5. Ahmad, A., Javaid, N., Alrajeh, N., Khan, Z.A., Qasim, U., Khan, A.: A modified
feature selection and artificial neural network-based day-ahead load forecasting
model for a smart grid. Appl. Sci. 5(4), 1756–1772 (2015)
Electricity Load Forecasting 603
6. Ahmad, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and
fast converging short-term load forecasting model for industrial applications in a
smart grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
7. Vrablecová, P., Ezzeddine, A.B., Rozinajová, V., Šárik, S., Sangaiah, A.K.: Smart
grid load forecasting using online support vector regression. Comput. Electr. Eng.
65, 102–117 (2018)
8. Liu, Y., Wang, W., Ghadimi, N.: Electricity load forecasting by an improved fore-
cast engine for building level consumers. Energy 139, 18–30 (2017)
9. Javaid, N., Hafeez, G., Iqbal, S., Alrajeh, N., Alabed, M.S., Guizani, M.: Energy
efficient integration of renewable energy sources in the smart grid for demand side
management. IEEE Access 6, 77077 (2018)
10. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics
and energy consumption forecasting. Energies 11(2), 452 (2018)
11. Abedinia, O., Amjady, N., Zareipour, H.: A new feature selection technique for load
and price forecast of electrical power systems. IEEE Trans. Power Syst. 32(1), 62–
74 (2017)
12. Javaid, N., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
I.A.: A hybrid genetic wind driven heuristic optimization algorithm for demand
side management in smart grid. Energies 10(3), 319 (2017)
13. Yu, C.N., Mirowski, P., Ho, T.K.: A sparse coding approach to household electricity
demand forecasting in smart grids. IEEE Trans. Smart Grid 8(2), 738–748 (2017)
14. González, J.P., San Roque, A.M., Pérez, E.A.: Forecasting functional time series
with a new Hilbertian ARMAX model: application to electricity price forecasting.
IEEE Trans. Power Syst. 33(1), 545–556 (2018)
15. Chitsaz, H., Zamani-Dehkordi, P., Zareipour, H., Parikh, P.P.: Electricity price
forecasting for operational scheduling of behind-the-meter storage systems. IEEE
Trans. Smart Grid 9(6), 6612–6622 (2018)
16. Moghaddass, R., Wang, J.: A hierarchical framework for smart grid anomaly detec-
tion using large-scale smart meter data. IEEE Trans. Smart Grid 9(6), 5820–5830
(2018)
17. Singh, N., Mohanty, S.R., Shukla, R.D.: Short term electricity price forecast based
on environmentally adapted generalized neuron. Energy 125, 127–139 (2017)
18. Luo, J., Hong, T., Fang, S.C.: Benchmarking robustness of load forecasting models
under data integrity attacks. Int. J. Forecast. 34(1), 89–104 (2018)
19. Bassamzadeh, N., Ghanem, R.: Multiscale stochastic prediction of electricity
demand in smart grids using Bayesian networks. Appl. Energy 193, 369–380 (2017)
A New System for Management of IoT
Sensors Considering Agile-Kanban
1 Introduction
Recently, many communication systems have been developed in order to support
humans. Especially, IoT technologies, IoT devices and sensors can provide many
services for humans. The IoT sensors can be embedded in electronic devices. For
instance, if a refrigerator has some IoT sensors inside, it can tell us the ingredi-
ents for cooking through the Internet anytime and anywhere. The temperature
of the air conditioner in the room can be controlled. Also, television, cooking
heater, microwave, bath-unit, toilet-unit, light and so on are going to provide
new services.
Currently, the vehicles can be connected to the Internet. Toyota uses the word
“Connected Car System”. This system can collect the data using mobile network
and sensors in the car. These are different kinds of data such as the data taken
by the car when is moving, car speed, braking or acceleration, weather condition
and traffic jam information. These data are gathered to the datacenter through
the network by using Internet technologies. The data can be analyzed and used
for solving some traffic problems. There are also other applications using IoT
such as the use of IoT technologies for agriculture, factories, offices, schools,
hospitals and so on.
Because the IoT devices and sensors generate a huge amount of packets, the
network load is increased and the sensor lifetime decreases. The lifetime of O2
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 604–611, 2019.
https://doi.org/10.1007/978-3-030-15035-8_58
A New System for Management of IoT Sensors Considering Agile-Kanban 605
2.1 IoT
Our life style is drastically changing by IoT technologies. The IoT can connect
various things to the Internet such as household electronics appliances, vehicles,
robots, communication devices, and some application softwares. Thus improving
our quality of life. In Fig. 1, we show the image of IoT environment. IoT sys-
tem has many sensors, which can be used for operating factories, farms, offices,
houses, and so on. However given the fact that most of IoT sensors are resource
limited and operate on batteries, the consumption power and life time of sensors
are important issues for wireless IoT sensor networks design [9,12].
There are some approaches for decreasing the number of packets in the net-
works. The Opportunistic Networks (OppNets) is one of them. It can provide
an alternative way to support the diffusion of information in special locations
in a city, particularly in crowded spaces where current wireless technologies can
exhibit congestion issues [8].
2.2 Kanban
Recently, Kanban system combined with Agile is used for software devel-
opment. Some papers use Kanban and Agile for collaboration work [3,10]. In
another paper, the authors use Agile approach with Kanban for managing the
security risk on e-commerce [2].
3 Proposed System
Kanboard is a free and open source software [4]. Kanboard user interface is
shown in Fig. 5. The Kanboard focus on simplicity minimalism. The number of
features is limited. By Kanboard can be known the current status of a project
because it is visual. It is very easy to understand and there is no need to explain
and no training is required.
• The system is able to analyze the IoT sensors status such as sensor’s working
time, life time, frequency of use and sensor’s activity.
• The proposed system can be installed in factory machines, vehicles, robots,
offices or school facilities, home electrical appliances and so on.
References
1. Abrahamsson, P., Salo, O., Ronkainen, J., Warsta, J.: Agile software development
methods: Review and analysis. arXiv preprint arXiv:1709.08439 (2017)
2. Dorca, V., Munteanu, R., Popescu, S., Chioreanu, A., Peleskei, C.: Agile approach
with kanban in information security risk management. In: 2016 IEEE International
Conference on Automation, Quality and Testing, Robotics (AQTR), pp. 1–6. IEEE
(2016)
3. Hofmann, C., Lauber, S., Haefner, B., Lanza, G.: Development of an agile develop-
ment method based on kanban for distributed part-time teams and an introduction
framework. Procedia Manuf. 23, 45–50 (2018)
4. Kanboard: Kanboard web page. https://kanboard.org/
5. Kiely, G., Kiely, J., Nolan, C.: Scaling agile methods to process improvement
projects: a global virtual team case study (2017)
6. Kirovska, N., Koceski, S.: Usage of kanban methodology at software development
teams. J. Appl. Econ. Bus. 3(3), 25–34 (2015)
7. Maneva, M., Koceska, N., Koceski, S.: Introduction of kanban methodology and
its usage in software development (2016)
8. Miralda, C., Donald, E., Kevin, B., Keita, M., Leonard, B.: A delay-aware fuzzy-
based system for selection of IOT devices in opportunistic networks considering
number of past encounters. In: Proceedings of the 21th International Conference
on Network-Based Information Systems (NBiS-2018), pp. 16–29 (2018)
9. Nair, K., Kulkarni, J., Warde, M., Dave, Z., Rawalgaonkar, V., Gore, G., Joshi,
J.: Optimizing power consumption in IOT based wireless sensor networks using
bluetooth low energy. In: 2015 International Conference on Green Computing and
Internet of Things (ICGCIoT), pp. 589–593. IEEE (2015)
10. Padmanabhan, V.: Functional strategy implementation-experimental study on
agile kanban. Sumedha J. Manag. 7(2), 6–17 (2018)
11. Petersen, J.: Mean web application development with agile kanban (2016)
12. Sheng, Z., Mahapatra, C., Zhu, C., Leung, V.: Recent advances in industrial wire-
less sensor networks towards efficient management in IOT. IEEE Access 3, 622–637
(2015)
13. Yacoub, M.K., Mostafa, M.A.A., Farid, A.B.: A new approach for distributed soft-
ware engineering teams based on kanban method for reducing dependency. JSW
11(12), 1231–1241 (2016)
A Two-Way Flow Model for Fog
Computing
1 Introduction
The IoT (Internet of Things) is composed of not only computers like servers
and clients but also devices like sensors and actuators installed in various things
[4]. Compared with traditional network systems like the cloud computing model
[1], the IoT is more scalable and huge amount of data from sensors are trans-
mitted in networks and are processed in servers. Networks and servers are too
heavily loaded to timely decide on actions for sensor data and deliver actions
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 612–620, 2019.
https://doi.org/10.1007/978-3-030-15035-8_59
A Two-Way Flow Model for Fog Computing 613
to actuators. In order to efficiently realize the IoT, the fog computing model is
proposed [7].
The fog computing model [7] is used to reduce the network and server traffic
of the IoT (Internet of Things). Since the IoT is scalable, huge amount of electric
energy is consumed in the IoT [2,3]. In order to not only increase the performance
but also reduce the electric energy consumption of the IoT, the TBFC (Tree-
based Fog Computing) model is proposed in our previous studies [4–6]. Here,
fog nodes are hierarchically structured in a height-balanced tree. A root node
shows a cloud composed of servers and leaf nodes are edge nodes which receive
sensor data from sensors and send actions to actuators. Each fog node has one
parent fog node and at least one child fog node. Each sensor first sends sensor
data to an edge fog node. The edge fog node is a parent of the sensor and the
sensor is a child of the edge fog node. Each fog node receives input data from
child fog nodes. Then, the fog node processes the data and sends the processed
data i.e. output data to a parent fog node. Finally, servers in a cloud receive
data processed by fog nodes. Then, the servers make a decision on actions to
be done by actuators and deliver the actions to actuators through fog nodes.
Thus, sensor data is sent upwards from sensors to a cloud of servers. Servers
in the cloud send actions downward to actuators. Here, it takes time to deliver
actions to actuators because servers make a decision on actions to be done by
the actuator.
In this paper, we newly propose a TWTBFC (Two-Way TBFC) model so
that actions to be performed on actuators are decided by fog nodes nearer to
devices in order to more promptly deliver actions to actuators. Here, each fog
node not only sends output data obtained by processing input data to a parent
fog node in a same way as the TBFC model but also forwards every input data
as well as the output data to the child fog nodes. Thus, each fog node can get
more global information by knowing output data sent by other sibling fog nodes
and the decision of the parent fog node. For example, a fog node of each house
can get not only data like energy consumption in the house but also neighbors
in a town. By using the data, each fog node can decide on actions to be executed
by descendant actuators and can send the actions downward to actuators. We
also present an application of a home energy control system in the TWTBFC
model.
In Sect. 2, we present a system model. In Sect. 3, we propose the TWTBFC
model. In Sect. 4, we present an application of the home energy control system
of the TWTBFC model.
2 System Model
2.1 Fog Computing Model
The fog computing (FC) model [7] to realize the IoT [4] is composed of devices,
fog nodes, and clouds. Clouds are composed of servers like the cloud computing
(CC) model [1]. Various devices, i.e. sensors and actuators are connected to edge
614 Y. Guo et al.
fog nodes. A sensor sends sensor data to edge fog nodes. Fog nodes are inter-
connected with other fog nodes in networks. An application process is composed
of subprocesses. A subprocess of an application process to handle sensor data is
performed on a fog node to process a collection of input data sent by sensors
and other fog nodes. Then, new output data, i.e. processed data is generated by
a fog node through performing the subprocess on the input data. For example,
a minimum value vk is selected by searching a collection of input data v1 , ...,
vl obtained from sensor s1 . The minimum value vk is the output data and a
collection of sensor data v1 , ..., vl are the input data of the fog node. Output
data processed by a fog node is sent to another fog node. Thus, sensor data
are processed by fog nodes and forwarded to fog nodes. Servers finally receive
data processed by fog nodes. Actions decided by servers are sent to edge fog
nodes and the edge fog nodes issue the actions to actuator nodes. On receipt of
an action, an edge fog node sends the action to actuators. Thus, subprocesses
are performed on not only servers but also fog nodes in the FC model while
centralized to servers in the cloud computing model.
Thus, the label R of a fog node fR shows a path from the root node f0 to the
fog node fR . For the length |R| of the label R of a fog node fR , the level of the
fog node fR is |R| − 1. For example, the levels of the fog nodes f0 and f011 are
|0| − 1 = 1 − 1 = 0 and |011| − 1 = 3 − 1 = 2, respectively. If the label R is a
concatenation R1 R2 of labels R1 and R2 , a fog node fR1 is an ancestor of the
fog node fR . For example, a fog node f021 is an ancestor of a fog node f02101
and a parent fog node of f01213 .
Let as(fR ) be a set of ancestor fog nodes of a fog node fR , i.e. {fR1 | R =
R1 R2 for some labels R2 }. Let ds(fR ) be a set of descendant fog nodes of a fog
node fR , {fR1 | R1 = RR2 for some labels R2 }. A least upper bound lub(fR , fS )
of a pair of fog nodes fR and fS is a common ancestor fog node fT whose any
descendant fog node is not an ancestor of the fog nodes fR and fS . SG(fR ) is a
set of fog nodes which are at the same level as a fog node fR .
A fog node fR takes input data dRi from each child fog node fRi (i = 0, 1,
..., lR − 1). DR shows a collection of input data dR0 , dR1 , ..., dR,lR −1 of the fog
node fR . A subprocess p(fR ) in the fog node fR does the computation on the
input data DR and then generates output data dR . Then, the fog node fR sends
the output data dR to a parent fog node pt(fR ). A collection DR of the input
data dR0 , dR1 , ..., dR,lR −1 and output data dR are stored by the storage module
SR in the storage DBR .
A notation |d| shows the size [Byte] of data d. The ratio |dR |/|DR | is the
output ratio ρR of a fog node fR . Suppose DR is a set {v0 , v1 , v2 , v3 } of four
numbers showing temperature obtained by child fog nodes fR0 , fR1 , fR2 , and
fR3 , respectively. If the output data dR is a maximum value v of the values
v0 , v1 , v2 , and v3 , the output ratio ρR of the fog node fR is |dR |/|DR | = 1/4.
616 Y. Guo et al.
Here, ρR ≤ 1 and the size of output data dR is smaller than the input data DR .
Suppose input data dRi from each fog node fRi is a set of values. If the output
data dR is obtained by taking the direct product of the input data dR0 , dR1 , ...,
dR,lR −1 the size |dR | of the output data dR is |dR0 | · |dR1 | · ... · |dR,lR −1 |. Here,
ρR ≥ 1.
A fog node fR is composed of input (IR ), computation (CR ), storage (SR ),
and output (OR ) modules as shown in Fig. 2 [6].
(i = 0, ..., alR − 1). Each edge fog node fR sends the actions acR0 , acR1 , ...,
acR,lR −1 to the child actuators aR0 , aR1 , ..., aR,alR −1 , respectively.
In the TWTBFC model, each fog node fR has two computation modules
CUR and CDR to process data U DR from child fog nodes fR0 , ..., fR,lR −1 and
to process data from the parent fog node pt(fR ).
4 Example
We consider an electric energy management system at home as an example of
the TWTBFC model. There are devices which supply electric energy to home
and consume electric energy at home. We consider the following five types of
devices:
1. Home (d0 ).
2. Rechargeable battery (RB) (d1 ).
3. Solar panel (SP) (d2 ).
4. Fuel cell (FC) (d3 ).
5. Electric company (EC) (d4 ).
Devices are classified in terms of two points, supplier and consumer of electric
energy. The home device d0 is just a consumer type, i.e. consumes electric energy.
The rechargeable battery (RB) device d1 is characterized by both consumer and
supplier type. The RB (d1 ) device supplies electric energy to the home device
(d0 ) but also consumes electric energy from the EC, SP, and FC devices. The
solar panel (SP) device d2 and fuel cell (FC) device d3 are devices of a supplier
type. The electric company (EC) device d4 is a supplier type, which supplies
electric energy not only to the home (d0 ) device but also the RB (d1 ) device.
For each device di , a pair of a sensor si and actuator ai are installed. Each
device di is connected to an edge fog node fi as shown in Fig. 5. Each sensor si
gives electric power [w] which is supplied to other devices and is consumed every
one second. By each actuator ai , the electric power supplied and consumed by
a device di is controlled.
A Two-Way Flow Model for Fog Computing 619
A node f0 is a root node which is a cloud servers. A fog node f0i exists for a
town (i = 0, ..., l0 − 1) and f0ij for each house (j = 0, ..., l0i − 1). Each edge fog
node f0ijk collects sensor data, i.e. electric power supplied and consumed five
devices d0 , d1 , d2 , d3 , d4 and sends the data to a parent fog node f0ij . Here, the
home device d0 just consumes electric energy supplied by the actuators a1 , ...,
a4 of the devices d0 , d1 , d2 , d3 , d4 . The battery (RB) device not only supplies
energy to the home device but also is supplied with energy from the solar panels
(SP), fuel cell (FC), and electric company (EC) devices.
5 Concluding Remarks
In this paper, we newly proposed the TWTBFC model to realize the IoT. In the
TBFC model, only a root fog node, i.e. server makes a decision on actions to
be executed by actuators. The actions are sent by the server to the actuators.
It takes time to deliver actions to actuators since edge nodes receive data from
sensors. In the TWTBFC model, fog nodes autonomously decide on actions by
using sensor data which are obtained by not only child fog nodes but also other
sibling fog nodes. Thus, each fog node can obtain more global data, i.e. data
from sibling nodes. We showed the electric power supply system as an example
of the TWTBFC model.
We are now evaluating the TWTBFC model.
620 Y. Guo et al.
References
1. Creeger, M.: Cloud computing: an overview. Queue 7(5), 3–4 (2009)
2. Enokido, T., Ailixier, A., Takizawa, M.: A model for reducing power consumption
in peer-to-peer systems. IEEE Syst. J. 4, 221–229 (2010)
3. Enokido, T., Ailixier, A., Takizawa, M.: An extended simple power consumption
model for selecting a server to perform computation type processes in digital ecosys-
tems. IEEE Trans. Ind. Inform. 10, 1627–1636 (2014)
4. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: An energy-
efficient model for fog computing in the internet of things (IoT). Internet Things
1–2, 14–26 (2018)
5. Oma, R., Nakamura, S., Duolikun, D., Enokido, T., Takizawa, M.: Evaluation of
an energy-efficient tree-based model of fog computing. In: Proceedings of the 21st
International Conference on Network-Based Information Systems (NBiS-2018), pp.
99–109 (2018)
6. Oma, R., Nakamura, S., Enokido, T., Takizawa, M.: A tree-based model of energy-
efficient fog computing systems in IoT. In: Proceedings of the 12th International
Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2018),
pp. 991–1001 (2018)
7. Rahmani, A.M., Liljeberg, P., Preden, J.-S., Jantsch, A.: Fog Computing in the
Internet of Things. Springer, Heidelberg (2018)
The 6th International Workshop on
Security Intricacies in Cyber-Physical
Systems and Services
(INTRICATE-SEC-2019)
Stochastic Methods to Find Maximum
Likelihood for Spam E-mail Classification
Abstract. The increasing volume of unsolicited bulk e-mails leads to the need
for reliable stochastic spam detection methods for the classification of the
received sequence of e-mails. When a sequence of emails is received by a
recipient during a time period, the spam filters have already classified them as
spam or not spam. Due to the dynamic nature of the spam, there might be emails
marked as not spam but are actually real spams and vice versa. For the sake of
security, it is important to be able to detect real spam emails. This paper utilizes
stochastic methods to refine the preliminary spam detection and to find maxi-
mum likelihood for spam e-mail classification. The method is based on the
Bayesian theorem, hidden Markov model (HMM), and the Viterbi algorithm.
1 Introduction
Spams are unwanted emails that the user does not want to have in his inbox. Email
Spam is one of the major problems of the today’s internet, bringing financial damage to
companies and annoying individual users. Spam is not only offensive and annoying; it
causes loss of productivity, decreases bandwidth and costs companies a lot of money.
Blocking spam email is considered a priority for network administrators and security
researchers. Spam filters are employed to assist the user in deciding if an email is worth
reading or not. There have been tremendous research efforts in this field that resulted in
a lot of commercial spam filtering products, such as: methods for construction of filters
to eliminate unwanted messages [1], comparison between the performances of machine
learning-based classifiers in filtering email spam [2, 3].
The challenging problem with spam filtering is the dynamic nature of the spam. The
characteristics (e.g. topics, frequent terms) of spam e-mail vary rapidly over time as
spammers always seek to invent new strategies to bypass spam filters [3]. One cannot
develop a filter and immediately implement it, because it will not have any basis for
classifying a document as spam or not spam. When a sequence of emails is received by
a recipient during a time period, the spam filters have already classified them as spam
or not spam. Due to the dynamic nature of the spam, there might be emails marked as
not spam but are actually real spams. It is important to be able to detect real spam
emails not only for personal convenience, but also for security. This paper proposes a
stochastic method to refine the preliminary spam detection. The method is based on the
Bayesian theorem, hidden Markov model (HMM), and the Viterbi algorithm.
In the spam detection ground work, the Bayesian filtering works by evaluating the
probability of different words appearing in legitimate and spam mails and then clas-
sifying them based on those probabilities [4, 5]. A hidden Markov model (HMM) is a
simple dynamic Bayesian network that is characterized by the initial state, state tran-
sition, and emission probabilities. This statistical model is derived by assuming that the
system under consideration is a Markov process with hidden states. The HMM is able
to correct deliberate misspellings, incorrect segmentations (adding/removing spaces),
and other word modifications [6–8].
The approach in this paper answers the following questions:
(a) How to find the probabilities (transition, initial, and emission) for the Hidden
Markov model?
(b) What is the initial probability that a random email is spam or not spam?
(c) If the indication (from the emails’ header) for a finite sequence of emails on a
finite time period, is marked as spam or not spam, what is the legitimate sequence
of emails, based on the stochastic model.
Given a sequence of states in different time frames (marked as spam and marked as not
spam), what are the most likely states that are congruent with the derived probabilities
and stochastic model based on training data. The training data in this case is the
confirmed and real emails identified as spam (S) or not spam (NS).
Suppose the training data is given by the sequence of emails:
S; S; S; S; NS; NS; S; S; S; S; NS; NS; S; S; S; ::; ::NS. . .; :. . .S. . .; . . .NS. . .; . . .S. . . ð1Þ
The transition probabilities are observed based on the number of times that the
switch from spam to not spam, not spam to spam happens and also the number of times
that the switch does not happen. These numbers are
PðSÞ ¼ 0:85 PðSÞ þ 0:35 PðNSÞ PðNSÞ ¼ 0:15 PðSÞ þ 0:65 PðNSÞ ð4Þ
Since P(S) + P(NS) = 1, P(S) = 0.7 and P(NS) = 0.3. Meaning that without any
knowledge or assumption from the header of emails, P(NS) (the probability that an
email is not spam) = 0.3 (a priori).
Stochastic Methods to Find Maximum Likelihood 625
The hidden states are marked as spam (MS) and marked as not spam (MNS). In
order to find emission probabilities, the Bayes’ theorem applies to conditional proba-
bilities calculation. The Bayes theorem states the prior probability (a priori): uncon-
ditional probabilities of our hypothesis before we get any data or any new evidence.
Simply speaking, it is the state of our knowledge before the data is observed. The
posteriori probability is a conditional probability about our hypothesis (our state of
knowledge) after we revised based on the new data [9].
For the hidden state MS, a posteriori, P(S|MS) = 0.85. Meaning that the probability
of the state to be spam is 0.85 if the hidden spam is marked as spam. For the hidden
state MNS, P(S|MNS) = 0.15. Meaning that if hidden state marked as not spam, then
the probability of the state to be spam is 0.15. Similarly, a posteriori, P(NS|MS) = 0.35
and P(NS|MNS) = 0.65.
Based on the training data, the emission probabilities emitted from the hidden states
are shown in Fig. 2.
Fig. 3. HMM
The Viterbi algorithm is a dynamic programming algorithm for finding the most likely
sequence of hidden states, the Viterbi path that results in a sequence of observed events
especially in the context of hidden Markov models [10]. For a simple case in which
there are two emails with the hidden states at time period Time 1 and Time 2, given
respectively by MS and MNS, what are the actual states based on HMM?
For the two states S and NS there are 22 = 4 transitions:
S!S S ! NS NS ! S NS ! NS ð5Þ
Based on the HMM Fig. 3, the probabilities associated with the four transition
states are as in Fig. 4.
Therefore, MLE for this case is 0.076. Meaning that the actual transition states
based on HMM is S ! S for MS and MNS.
For the three emails with the hidden states at time period Time 1, Time 2, and Time
3, given respectively by MS, MNS, and MS, based on HMM, the actual states are
calculated as follows.
For the two states S and NS there are 23 = 8 transitions:
S ! S ! S NS ! S ! NS S ! NS ! S NS ! NS ! NS ð6Þ
NS ! S ! S S ! S ! NS NS ! NS ! S S ! NS ! NS ð7Þ
Based on the HMM, comparing the probabilities associated with the eight transition
states, the MLE for this case is MLE = 0.7 * 0.85 * 0.85 * 0.15 * 0.85 * 0.85 = 0.055.
Meaning that the actual transition states based on HMM is S ! S ! S, for MS, MNS,
and MS.
Stochastic Methods to Find Maximum Likelihood 627
With the same token for the case of a sequence with 4 time periods, there are
24 = 16 paths that are required to be checked for MLE. For 5 time periods, there are
25 = 32 paths and for 6 time periods there will be 26 = 64 paths. It is seen that the
number of paths increases exponentially.
In order to reduce the number of paths from 32, 64, and … to only one path, the
dynamical programming and specifically Viterbi algorithm is the natural approach to
achieve the solution.
Consider the sequence of hidden states MS, MS, MNS, MNS, MNS, MS given for
the time periods Time 1, 2, 3, 4, 5, 6 respectively. By utilizing the Viterbi algorithm,
based on HMM (Fig. 3), the MLE probabilities for the S and NS states are:
For the time period Time 1, between the two probabilities
The MLE is 0.595. Therefore the state S is the first candidate at Time 1.
For the time period Time 2, between the four probabilities
PðSÞ ¼ 0:595 0:85 0:85 ¼ 0:429 PðSÞ ¼ 0:105 0:35 0:85 ¼ 0:031 ð9Þ
PðNSÞ ¼ 0:595 0:15 0:35 ¼ 0:031 PðNSÞ ¼ 0:105 0:65 0:35 ¼ 0:024
ð10Þ
The MLE is 0.429. Therefore the state S is the next candidate at Time 2.
For the time period Time 3, between the four probabilities
PðSÞ ¼ 0:429 0:85 0:15 ¼ 0:055 PðSÞ ¼ 0:031 0:35 0:15 ¼ 0:002 ð11Þ
PðNSÞ ¼ 0:429 0:15 0:65 ¼ 0:042 PðNSÞ ¼ 0:031 0:65 0:65 ¼ 0:013
ð12Þ
The MLE is 0.055. Therefore the state S is the next candidate at Time 3.
Up to here the sequence is S ! S ! S.
For the time period Time 4, between the four probabilities
PðSÞ ¼ 0:055 0:85 0:15 ¼ 0:007 PðSÞ ¼ 0:042 0:35 0:15 ¼ 0:002
ð13Þ
PðNSÞ ¼ 0:055 0:15 0:65 ¼ 0:005 PðNSÞ ¼ 0:042 0:65 0:65 ¼ 0:018
ð14Þ
PðSÞ ¼ 0:007 0:85 0:15 ¼ 0:0009 PðSÞ ¼ 0:018 0:35 0:15 ¼ 0:001
ð15Þ
PðNSÞ ¼ 0:007 0:15 0:65 ¼ 0:0007 PðNSÞ ¼ 0:018 0:65 0:65 ¼ 0:008
ð16Þ
Fig. 5. For the sequence MS, MS, MNS, MNS, MNS, MS the real states are:
S ! S ! S ! NS ! NS ! S
PðSÞ ¼ 0:001 0:85 0:85 ¼ 0:0007 PðSÞ ¼ 0:008 0:35 0:85 ¼ 0:0024
ð17Þ
PðNSÞ ¼ 0:001 0:15 0:35 ¼ 0:00005 PðNSÞ ¼ 0:008 0:65 0:35 ¼ 0:0018
ð18Þ
The MLE is 0.0024, the state S is the candidate at Time 6 (Fig. 5).
At this point the sequence is S ! S ! S ! NS ! NS ! S.
The concept of conditional probability and Bayesian theory is essential and pro-
found [11]
primary = []
if finals[0] == 'S':
probabilities.append((p_s*p_sh, p_r*p_rh))
else:
probabilities.append((p_s*p_sg, p_r*p_rg))
for i in range(1,len(finals)):
yesterday_sunny, yesterday_rainy = probabilities[-1]
if finals[i] == 'A':
today_sunny = max(yesterday_sunny*p_ss*p_sh,
yesterday_rainy*p_rs*p_sh)
today_rainy = max(yesterday_sunny*p_sr*p_rh,
yesterday_rainy*p_rr*p_rh)
probabilities.append((today_sunny, today_rainy))
else:
today_sunny = max(yesterday_sunny*p_ss*p_sg,
yesterday_rainy*p_rs*p_sg)
today_rainy = max(yesterday_sunny*p_sr*p_rg,
yesterday_rainy*p_rr*p_rg)
probabilities.append((today_sunny, today_rainy))
print(probabilities)
for p in probabilities:
if p[0] > p[1]:
primary.append('MS')
else:
primary.append('MNS')
print (primary)
print(finals)
For the marked sequence [‘MS’, ‘MS’, ‘MNS’, ‘MS’, ‘MNS’, ‘MS’, ‘MS’, ‘MS’,
‘MS’, ‘MS’, ‘MS’] as an input, by executing the above python program, the output is:
[‘S’, ‘NS’, ‘NS’, ‘S’, ‘NS’, ‘S’, ‘S’, ‘S’, ‘NS’, ‘S’, ‘NS’].
The method in this paper can be utilized to address the following future works:
1 – Speech recognition
2 – Genetics DNA sequences
3 – Vehicle location
4 – Text tagging WORDS to TAGS (the definite articles, preposition, verb, noun).
References
1. Sahami, M., Dumais, S., Heckerman, D., Horvitz, E.: A Bayesian approach to filtering junk
e-mail. In: AAAI 1998 Workshop on Learning for Text Categorization (1998)
2. Al-Jarrah, O., Khaterz, I., Al-Duwairi, B.: Identifying potentially useful email header
features for email spam filtering. In: ICDS The Sixth International Conference on Digital
Society (2012)
632 S. M.-H. Mansourbeigi
3. Awad, W.A., ELseuofi, S.M.: Machine learning methods for spam e-mail classification. Int.
J. Comput. Sci. Inform. Technol. 3(1), 173–184 (2011)
4. Eberhardt, J.J.: Bayesian spam detection. Scholarly Horizons: University of Minnesota,
Morris Journal, vol. 2, no. 1 (2015)
5. Freeman, D.M.: Using naive Bayes to detect spammy names in social networks. In:
Proceedings of the 2013 ACM Workshop on Artificial Intelligence and Security, AISec
2013, pp. 3–12. New York (2013)
6. Lee, S., Jeong, I., Choi, S.: Dynamically weighted hidden Markov model for spam
deobfuscation. In: IJCAI 2007, Proceedings of the 20th International Joint Conference on
Artificial Intelligence, pp. 2523–2529 (2007)
7. Lee, H., Ng, A.Y.: Spam deobfuscation using a hidden Markov model. In: Proceedings of
the 2nd Conference on Email and Anti-Spam, Stanford University (2005)
8. Jaswal, V., Sood, N.: Spam detection system using hidden Markov model. Int. J. Adv. Res.
Comput. Sci. Softw. Eng. 3(7), 304–308 (2013)
9. Roy, S., Patra, A., Sau, S., Mandal, K., Kunar, S.: An efficient spam filtering techniques for
email account. Am. J. Eng. Res. 2(10), 63–73 (2013)
10. Viterbi, A.J.: Error bounds for convolutional codes and an asymptotically optimum decoding
algorithm. IEEE Trans. Inf. Theor. 13(2), 260–267 (1967)
11. Papoulis, A.: Probability, Random Variables, and Stochastic Processes. McGraw-Hill Series
in Electrical Engineering, 3rd edn. (1991)
Design and Research of Trusted Acquisition
Terminals Based on Domestic Password
1 Introduction
With the gradual development of smart grid in China, the construction of power
consumption information collection system is becoming more and more perfect. Fre-
quent data interaction of power consumption information collection system brings
security risks [1]. According to statistics, most of the security attacks are initiated by
terminal devices, which are vulnerable to various types of security threats and attacks,
such as denial of service attacks, eavesdropping, illegal intrusions, and so on. These
risks will lead to the leakage of sensitive information and the control of the terminal by
hackers, thus threatening the normal operation of the whole system [2, 3]. Although it
is one of the security weaknesses in the system, the security of the terminal system has
not received enough attention.
The concept of trusted computing makes people transfer the idea of solving
information security problems to solving terminal security problems [4]. At present, the
research in the field of trusted computing in China has gone from the initial tracking
learning to the stage of independent innovation. At the same time, China attaches great
importance to information security, and chinese codes have been developing rapidly.
The application of domestic codes in trusted computing can effectively reverses the
current situation that the security protection technology of key information systems in
our country relies heavily on foreign countries [5], which can improve the autonomous
and controllable ability in the field of security technology in our country.
Therefore, this paper combines the current status of power consumption informa-
tion collection system, applies the trusted computing technology based on domestic
cryptographic algorithm to the security reinforcement of terminal equipment, and puts
forward the security design scheme of the acquisition terminal, which can effectively
improve the security of the acquisition terminal, and have urgent realistic needs and
research significance.
I/O
Port
adds symmetric cryptographic algorithm, which reduces the complexity of key man-
agement, data encryption and authorization protocol. The SMS4 is used as a symmetric
cipher algorithm.
the acquisition terminal, the acquisition terminal can not be effectively monitored,
physical protection can not effectively prevent attacks; there is still a certain gap
between the security hardware encryption capability and the data requirements of the
acquisition terminal; for the operation system and application program which under-
takes the main work of the acquisition terminal, there is still no good detection means;
the acquisition terminal currently adopts modular design, even if the functional mod-
ules are safe, there will still be security risks in the combination of the modules [13].
and cooperates with the trusted acquisition terminal operating system to realize the
integrity, identity trustworthiness and data security protection of the acquisition terminal.
Trusted acquisition terminal architecture mainly includes three parts: trusted
hardware system, trusted operating system and trusted application program. In the
hardware system of the trusted acquisition terminal, TCM communicates with the main
control unit module of the acquisition terminal through the trusted module interface
circuit (TMIC). TMIC also has the function of assisting TCM to complete the trusted
verification and start-up control of Bootloader, which will be introduced in detail later.
Trusted operating system mainly includes software layer and service layer. Based on
TCM supporting software TDDL, TCS and TSP, it completes three functions: trusted
measurement, trusted message and trusted storage, realizes integrity verification,
identity authentication and security protection of acquisition terminal, and provides
trusted service for acquisition terminal application.
MCU
ARM7
Input Communication
/Output Interface Circuit
Data Memory
Circuit
Program
Memory
Trusted Module
Interface Circuit
Clock Circuit
Auxiliary
Other Circuit Speech Control
Circuits Circuit
CPU
Basic Circuit
Trusted module interface circuit (TMIC) is the key to the realization of trusted
acquisition terminal system. Besides providing communication channel between TCM
and MCU, it also has the function of assisting TCM to complete trusted boot loader.
The specific process of trusted start-up of trusted acquisition terminal system is as
follows:
(1) TCM starts and checks itself first after the system is powered up;
(2) TMIC reads Bootloader to TCM and verify its integrity;
(3) If the integrity checks pass, TMIC establishes the actual connection between
MCU and Bootloader, TCM sends license signal to MCU processor, and Boot-
loader starts; if the check is wrong, reset the system.
TCM self-check
TMIC error
bootloader
check
Success TMIC
MCU
Bootloader start
operating system
Application
Step 5. Bootloader calls TCM’s interface command, calculates the hash value of
operating system, and completes the integrity check by comparing with the baseline
value. If the check passes, Bootloader transfers control to the operating system and
enters step 6, otherwise it restores and reports errors.
Step 6. The operating system calculates the hash value of the application and
completes the integrity check by comparing with the baseline value. After the check
is passed, the control will be given to the application and enter step 7. Otherwise,
the application will be restored and the error will be reported.
Step 7. application goes into operation. By measuring step by step and transferring
trust chain in such a trusted way, we can verify the security and trustworthiness of
each application, and fundamentally protect the security of the whole collection
terminal system.
The gradual informationization and interaction of smart grid make the security of
power information acquisition terminal more and more important. In recent years, the
domestic cryptography technology and trusted technology have developed rapidly in
our country. In this paper, the trusted computing technology based on the domestic
cryptography is applied to the security reinforcement of the terminal of power infor-
mation acquisition system. The design of trusted acquisition terminal is studied from
three aspects: overall architecture, trusted start-up and trust chain transmission, which
640 Y. Feng et al.
can effectively enhance the safety protection ability of acquisition terminals and bring
new ideas for the safety reinforcement of the acquisition terminals.
References
1. Zhang, S.M., Wang, Z.Y., Wang, B.Y.: Terminal integrity detection scheme of electricity
information acquisition system based on trusted computing. Electr. Power Autom. Equip. 37
(12), 60–66 (2017)
2. The Research and Implementation of Embedded Power Distribution Terminal Based on
Trusted Computing. North China Electric Power University, Beijing
3. Shen, C.X., Zhanu, H.U., Fend, D.U., et al.: Survey on information security. Sci. China Ser.
E Inf. Sci. 37(2), 129–150 (2007)
4. Department of Defense Computer Security Center: Department of Defense Trusted
Computer System Evaluation Criteria. DOD 5200.28-STD. DOD, USA (1985)
5. Shen, C.X., Gong, B.: The innovation of trusted computing based on the domestic
cryptography. J. Cryptol. Res. 2(5), 381–389 (2015)
6. Trusted Computing Group. TCG [OL]. https://www.trusted.computinggroup.org
7. Trusted Computing Group [OL]. https://www.trusted.computinggroup.org/specs/
8. Trusted Computing Group, TCG Specification Architecture Overview [EB/OL]. Revision
1.4 (2007)
9. Trusted Mobile Platform Protocol Specification Document [EB/OL]. http://www.trusted-
mobile.org/TMP_Protoeol_revl_00.Pdf
10. Yang, B., Wu, Z.-Q., Fu, X.-P.: Dynamic integrity measurement model based on trusted
computing. Comput. Eng. 38(2) (2012)
11. Yu, Z., Zhang, W., Dai, H., et al.: A trusted architecture for virtual servers with trusted
platform module and certify-cafe authority. J. Signal Process. Syst. 86(2/3), 327–336 (2017)
12. Mukhamedov, A., Gordon, A.D., Ryan, M.: Towards a verified reference implementation of
a trusted platform module. In: Security Protocols XVII, pp. 69–81. Springer, Heidelberg
(2013)
13. Zhai, B., Ni, M., Shi, Y., Fan, P.: A review on the security of embedded system. J. Wuhan
Univ. (Nat. Sci. Ed.) 64(2), 95–108 (2018)
14. Xu, M.D., Zhang, H.G., Zhang, F., et al.: Survey on chain of trust of trusted system. Acta
Electron. Sin. 42(10), 2024–2031 (2014)
15. Chang, D.X., Feng, D.G., Qin, Y., et al.: Analyzing the trust chain of trusted virtualization
platform based on the extended LS2. J. Commun. 34(5), 31–41 (2013)
Trust-Based Security Mechanism for Detecting
Clusters of Fake Users in Social Networks
Abstract. Inspired by the ability of the trust mechanism which can capture
human trust using measurement theory we have designed trust-based security
mechanism to find clusters of fake users in social networks. Proposed framework
enables us to take human and community knowledge into a loop by using a
variation of trust to detect fake users. This framework will increase the accuracy
of detecting fake users by using both human and machine knowledge. Experi-
ments are performed on randomly generated social networks to validate the
potential of this framework. Results show that the variation of trust over time
can able to differentiate between clusters of fake users from clusters of real users.
1 Introduction
With the rise in the use of technology, the use of social media is also increasing. Lots of
people use social media to share pictures, videos, news and to stay up to date with the
current events and news. People are also using social networks to connect with other
people. However, with so many advantages of social media, there are some disad-
vantages also. Bad people are misusing social media to spread fake news, malicious
content and viruses by creating fake user accounts, that cannot easily be caught. So, it
is imperative to detect these fake user accounts as soon as possible to minimize the
harm caused by them.
To detect fake user accounts, it is essential to understand how social networks
work. All our social interactions whether they are face to face or over the Internet are
based on the trust which is based on the number of positive and negative interactions
happened between the people and is developed over time and can be used to detect fake
users. [1] In this paper a trust-based mechanism is proposed for social networks to filter
out malicious content and to study the damage caused by the users based on their trust
values. This type of mechanism is helpful in studying the behavior of fake users.
We have developed a method to detect fake accounts on the social network based
on a trust-based security mechanism. In this method, we have integrated trust-based
mechanism with machine learning to detect clusters of fake accounts. The trust-based
mechanism is used to capture human interactions with each other, and machine
learning is used to process and analyze a large amount of social network data. Trust-
based mechanism is used to map human interactions into physical measurements. This
mechanism will help to capture the human knowledge about other humans or events
which can be very beneficial for detecting fake users. Machine learning algorithms are
used to capture characteristics of the social networks. It is used to cluster the social
network graph based on their characteristics. These clusters are then analyzed based on
the trust values, density, the variation of the trust values over time to label them as a
cluster of legitimate users or cluster of fake users. The proposed method of detection
can be beneficial as it can capture the community information to detect fake users.
This paper is organized as follows. In Sect. 2 we present the background and
related work in the fields of fake user detection and trust mechanism. Section 3
describes our framework of Fake user detection. In Sect. 4 we present different
experiments and simulations results of our framework. Also, in Sect. 5, we conclude
the paper.
This section gives information about the background and related work done in the field
of trust and security.
There are several approaches where the community information is used in detecting
fake users [1, 5, 8]. Some of the approaches used security experts to manually check
user profiles which the system ranked them as more likely to be fake to keep human
knowledge in the loop for detecting fake users [14]. Some of the approaches used trust
framework to calculate the damage caused by attackers based on their trust value
assigned by the local community [1].
In our approach, we have used the trust mechanism to capture local and community
knowledge used to detect fake clusters.
To make fake user detection more efficient, we have introduced Trust-based Security
Mechanism. This approach considers human and community knowledge along with
machine learning to detect clusters of fake users. We have designed evaluation metrics
based on trust between the nodes of the clusters and using features of social network
graph to label detected clusters as the clusters of fake users or the clusters of good
users.
This algorithm consists of three steps. The first step is to calculate the trust of the
users of the social network using their interactions with each other. Secondly is to
644 D. Kaur et al.
detect clusters in social network graph using a machine learning algorithm. Last is to
analyze clusters based on trust to label them as good or bad clusters.
Section 3.1 presents Trust framework used to calculate trust between the users of the
social network. Section 3.2 explains the clustering algorithm we have used to cluster
social network graph Sect. 3.3 explains the evaluation metrics used to label the clusters.
Both impressions and confidence are used to calculate the value of trust. We
normalized the values of impression and confidence between 0 and 1, but it can be
varied based on the circumstances. These values are used to represent trust between
users of the social network.
However, if there are no interactions happened between the nodes, then trust is
calculated using indirect links. Trust framework has two methods to calculate indirect
trust between the nodes, that are Trust Transitivity and Trust Aggregation.
Trust Transitivity: This method is used to calculate indirect trust. If a person X does
not have a direct connection to Z, but X knows someone for example Y who knows Z,
then trust from X to Z is calculated through Y using trust transitivity. So, the trust
between X and Z will be concatenation between trust from X to Y and Y to Z.
Equation 3 and four show how impression and confidence are calculated for transitive
trust.
Z ¼m
mXY mYZ ð3Þ
XY
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ¼
vXY ðvXY Þ2 ðmYZ Þ2 þ ðvYZ Þ2 ðmXY Þ2 ð4Þ
Trust Aggregation: This method is used to summarize trust value if there are more
than one indirect path between nodes. For example, to reach D, there are two paths A-
B-D and A-C-D from A then the aggregated trust will be the combination of both the
trust values of the graph. Equation 5 and six show how impression and confidence are
calculated for trust aggregation.
mADB þ mC
AD
B mC ¼
mAD ð5Þ
AD
2
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 AD 2
ðvADB Þ þ vC
B vC
vAD ¼ ð6Þ
AD
22
Using this trust framework, we can calculate the trust of every user based on his/her
interactions with other users on social networks.
To divide the social network graph into clusters, the spectral clustering method is
used. This clustering method has excellent performance in graph-based clustering and
is a combination of k-means clustering and DBSCAN.
The main aim of clustering the social network graph is to divide the graph into
groups of highly correlated subgraphs that share similar properties. That is, to maxi-
mize intracluster similarity and minimize intercluster similarity. This is performed to
deal with a large amount of social network data, as it would be more efficient to apply
the trust mechanism on graphs with similar properties.
After these clusters are detected, they are analyzed based on trust mechanism and
other features to label them as the clusters of good users or the cluster of fake users.
Machine learning algorithms enable us to cluster a large amount of data efficiently.
This makes the process of detecting fake users very fast. In the following section, we
will discuss the evaluation metrics based on which fake users are detected.
There can be a case where the density of the cluster of real users is also high, in that
case, the trust of the cluster and variation of trust over time makes the difference. So,
the density of the cluster is used along with trust to detect fake users.
Cluster formation based on time: Time taken by the fake users to build connections
within themselves is very less than the time taken by the real users to build actual
connections. Both the density of the cluster and time taken to form the clusters are
interrelated. Clusters or communities formed by fake users are within themselves, and
fake user profiles are only created if they need to perform some malicious activity, so
these clusters are formed instantly while the clusters of the real users take a significant
amount of time to stabilize and form communities. These features imply that clusters
Trust-Based Security Mechanism for Detecting Clusters of Fake Users 647
formed at the instant of time are more probable of being clusters of fake users then the
clusters formed in a significant amount of time. Using this feature of time along with
trust feature enable us to detect clusters of fake users.
Variation of trust over time: Trust is the essential feature to detect fake users. Trust
is used to capture local information about the users. Trust-based metrics are used to
differentiate real users from fake users. These metrics are divided into two parts: First is
based on the average trust of the clusters and Second is based on the variation of trust
over time.
Fake user clusters tend to have a high trust value as compared to real user clusters.
This is based on how fake users make connections with other fake users. For spreading
malicious content, fake users need to increase their trustworthiness among other users
because if their trust is low, they can easily be detected as fake users. They do so by
exchanging positive messages with other fake users to increase their trust instantly.
Trust of fake user tend to have fewer fluctuations, and it will mostly increase with
time, while the trust of real users will have move fluctuations and it will go more up
and down based of the positive and negative interactions happened between them.
So, by analyzing the trust of the clusters and how the trust changes with time along
with the properties of density and time, fake user clusters can be detected.
4 Implementation
4.1 Dataset
As the fake user data of social networks is not available publicly, we generated syn-
thetic social network graph that can mimic the properties of the real graph to experi-
ment our approach.
To simulate the social network graph, we have used the Erdos-Renyi model to
generate the random social graph. This type of model is very useful in generating
random social graphs [15]. If G (V, E) is the random graph generated, then V represents
users of the social network and E represents the relationship between users, which is
chosen randomly with the probability p such that p belongs to set [0, 1]. The degree of
these nodes is also random numbers selected from the binomial distribution [15].
Because of the randomness feature of this graph model, it is perfect for generating
random social graphs.
We have created a social graph of 1000 nodes, with less probability of edge
creation. This created graph mimics the social network graph of good users as the
nodes in this graph is distributed and sparsely connected. To mimic fake user behavior,
we increased the probability of edge creation so that the created graph is more densely
connected. Timestamp at which edge is created is also randomly generated.
To capture the trust values between the nodes, values of impression and confidence
are randomly generated. This generation of trust is divided into two sections. In one
section trust values is randomly assigned between [0, 1] for sparsely connected
648 D. Kaur et al.
distributed graph and for densely connected graph higher trust values are assigned
between [0.5, 1] to mimic the behavior of fake users.
All the data generated by the above methods are used to test our approach of
detecting fake user clusters.
Fig. 1. Three clusters detected from social network graph which have the most similar
properties within the cluster and most different properties with another cluster.
Results of the clustering and analysis of the properties show that the density of
cluster 3 is very high as compared to the density of cluster 1 and cluster 2. Also, the
average trust value of cluster 3 is also high. This shows that the nodes in cluster 3 have
high trust connections within the densely connected small community and have
decidedly fewer connections to outside communities. So, based on trust, the probability
of users in cluster 3 being fake is high as compared to cluster 1 and two because
gaining high trust only from a small community in less time is very rare.
Trust-Based Security Mechanism for Detecting Clusters of Fake Users 649
To further validate our point, we have used an evaluation metric based on density,
time and trust to categorize the detected clusters. Cluster 1 is sparsely connected
average trust cluster of users, present on the network for more than a year. Cluster 2 is
also a sparsely connected average trust cluster created at the instant of time. Cluster 3 is
densely created high trust cluster created at the instant of time. All these detected
clusters can either be clusters of real users or clusters of fake users.
To show how variation of trust can able to detect fake clusters accurately, we have
plotted six months trust data of the clusters.
1
0.9
0.8
0.7
0.6
Trust
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6
Months
BadCluster GoodCluster(Old) GoodCluster(New)
Fig. 2. Change of trust values over time for different clusters. Clusters with high trust value
during creation and fewer fluctuations over time are more probable of being clusters of fake users.
Above Fig. 2 shows the graph of trust values of the clusters over time. The graph
validates our point of fake users being gaining very high trust in the creation of the
cluster and maintaining it over time while real users tend to gain trust more gradually
and have more fluctuations.
5 Conclusion
Acknowledgments. This work was partially supported by the National Science Foundation
under Grant No. 1547411 and by the U.S. Department of Agriculture (USDA) National Institute
of Food and Agriculture (NIFA) (Award Number 2017-67003-26057) via an interagency part-
nership between USDA-NIFA and the National Science Foundation (NSF) on the research
program Innovations at the Nexus of Food, Energy and Water Systems.
References
1. Zhang, P., Durresi, A.: Trust management framework for social networks. In: 2012 IEEE
International Conference on Communications (ICC), Ottawa, ON, pp. 1042–1047 (2012)
2. Zhang, P., Durresi, A., Barolli, L.: Survey of trust management on various networks. In:
2011 International Conference on Complex, Intelligent, and Software Intensive Systems,
Seoul, pp. 219–226 (2011)
3. Ruan, Y., Durresi, A., Uslu, S.: Trust assessment for internet of things in multi-access edge
computing. In: 2018 IEEE 32nd International Conference on Advanced Information
Networking and Applications (AINA), Krakow, pp. 1155–1161 (2018)
4. Ruan, Y., Durresi, A., Alfantoukh, L.: Using Twitter trust network for stock market analysis.
Knowl.-Based Syst. 145, 207–218 (2018)
5. Ruan, Y., Zhang, P., Alfantoukh, L., Durresi, A.: Measurement theory-based trust
management framework for online social communities. ACM Trans. Internet Technol. 17
(2), 24 (2017). Article 16
6. Ruan, Y., Durresi, A.: A survey of trust management systems for online social communities
– trust modeling, trust inference and attacks. Knowl.-Based Syst. 106, 150–163 (2016)
7. Ruan, Y., Durresi, A., Alfantoukh, L.: Trust management framework for internet of things.
In: 2016 IEEE 30th International Conference on Advanced Information Networking and
Applications (AINA), Crans-Montana, pp. 1013–1019 (2016)
8. Cao, Q., Sirivianos, M., Yang, X., Pregueiro, T.: Aiding the detection of fake accounts in
large scale social online services. In: Proceedings of the 9th USENIX Conference on
Networked Systems Design and Implementation (NSDI 2012), p. 15. USENIX Association,
Berkeley (2012)
9. McAuley, J., Leskovec, J.: Learning to discover social circles in ego networks. In: Pereira,
F., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.) Proceedings of the 25th International
Conference on Neural Information Processing Systems - Volume 1, (NIPS 2012), vol. 1,
pp. 539–547. Curran Associates Inc. (2012)
10. Fire, M., Katz, G., Elovici, Y.: Strangers intrusion detection-detecting spammers and fake
profiles in social networks based on topology anomalies. Hum. J. 1(1), 26–39 (2012)
11. Ferrara, E., Varol, O., Davis, C., Menczer, F., Flammini, A.: The rise of social bots.
Commun. ACM 59(7), 96–104 (2016)
12. Xiao, C., Freeman, D.M., Hwa, T.: Detecting clusters of fake accounts in online social
networks. In: Proceedings of the 8th ACM Workshop on Artificial Intelligence and Security.
ACM (2015)
13. Zhang, Z., Kulkarni, S.R.: Detection of shilling attacks in recommender systems via spectral
clustering. In: 2014 17th International Conference on Information Fusion (FUSION). IEEE
(2014)
14. Kagan, D., Elovichi, Y., Fire, M.: Generic anomalous vertices detection utilizing a link
prediction algorithm. Soc. Netw. Anal. Min. 8(1), 27 (2018)
15. Hamlili, A.: Intelligibility of Erdös-Rényi random graphs and time varying social network
modeling. In: ICSDE (2017)
Machine Learning Based Approach
to Detect Wormhole Attack in VANETs
1 Introduction
Vehicular ad-hoc networks (VANETs) is a subclass of mobile ad-hoc net-
works (MANET), with characteristics of high mobility, frequently changing net-
work topology and additional complexities [5]. VANETs facilitate many safety
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 651–661, 2019.
https://doi.org/10.1007/978-3-030-15035-8_63
652 P. K. Singh et al.
2 Related Work
In 2003, the Hu et al. [11] introduced the wormhole attack and discussed how
severe this attack is in mobile adhoc networks. It is challenging to defend against
and can form a serious threat especially against many adhoc routing protocols.
In [4,20], the authors provided the details of security threats to adhoc routing
protocols and highlighted that the wormhole attack is one of the most powerful
attacks. As a result, various techniques have been proposed to detect, avoid and
defend the wormhole attacks. The detection techniques are categorized into three
categories [1]: hardware-based, anomaly-based, and guard node-based mecha-
nisms.
There are some more methods exist in the literature to detect wormhole
attacks in VANETs. However, most of these proposals either have issues in terms
of communication overhead and delay in detection, or they require time synchro-
nization, localization, and special hardware to achieve the goal. We believe that
ML can play an important role here to address the issues of the existing system
and can help detect with better accuracy without incurring additional overhead
and complexity.
Machine Learning Based Approach to Detect Wormhole Attack in VANETs 653
– Each node keeps track of its neighbors by listening hello messages that each
node broadcast at defined interval.
– If a node wants to send data to the node which is not its neighbor, then it
broadcast RREQ (Route Request) message.
– Every neighbor which receives RREQ, either return RREP (Route Reply) (if
know the route to destination or destination itself) or rebroadcast RREQ.
– When destination is found, then RREP is sent via the reverse path.
– A higher sequence number signifies a fresh route, and the forward path is
used to route data packets.
– The RERR (Route Error) message allows AODV to adjust routes when the
node moves around.
The wormhole attack can be launched by a malicious node, which may be either
an outsider node or an insider node. The outsider node does not have crypto-
graphic materials such as certificate and keys are given by the CA whereas the
insider possesses the required certificates and keys. In this work, we consider
internal adversaries who in real life may be created by compromising legitimate
nodes. These compromised nodes can have Byzantine behavior and can also
cooperate or collude amongst themselves.
The adversaries launches wormhole attacks by exploiting the vulnerabilities
of popular ad-hoc routing protocols, such as DSR and AODV. The common
methods of launching wormhole attack are [16]: packet encapsulation, out-of-
band channel (OBC), high power transmission, packet relay and protocol devi-
ations. In this work, we consider wormhole using packet encapsulation. Due to
space constraint we are not providing the details of this attack. However, authors
can refer works [1,16] for details of all these attacks.
The wormhole attack has been considered as one of the severe security
attacks, which can severely affect the multihop communication in VANETs. The
malicious nodes who collude to from the tunnel, attract routes through them and
use it to pass the routing traffic. Variety of attacks can be launched on the data
traffic flowing through the wormhole tunnel, such as dropping the data packets
selectively. This attack can also prevent nodes to discover legitimate routes and
thus disrupt normal functionality of VANETs. Moreover, it can also affect data
aggregation, location-based wireless security, and clustering protocols [16].
4 Methodology
In our work, we propose a ML-based wormhole attack detection system that
determines the behavior of vehicles in VANETs. Our system uses the trace files
produced by simulator that consists both normal and abnormal behavior (under
wormhole attack) of nodes in VANETs. The steps below provide the details of
methodology adopted for our proposed system:
656 P. K. Singh et al.
4.1 Simulation
We used two simulators to model the VANET scenario discussed in the previous
section: Simulation of Urban Mobility Model (SUMO) [6] as a traffic simula-
tor and NS3 as a network simulator [10]. We used the Manhattan mobility by
importing a map from OpenStreetMap (OSM) and converting to usable NS3 file
by generating trace using NETCONVERT, POLYCONVERT, randomTrips, and
traceExporter features of the SUMO. Figure 3a depicts the steps to produce the
SUMO configuration file from an OSM file [21]. Figure 3b shows the OSM file
and corresponding SUMO net file used in work. The selected road segment is of
size 3 km2 and each vehicle is of length 5 meters and max speed of 15 m/s.
The mobility trace file generated by the SUMO is fed into the NS3 simulator
using its ns2mobilityhelper module. The Table 1 shows the parameters that we
used for our simulation scenario.
Parameter Value
Simulator NS 3.26
Simulation time 400 s
Number of vehicles 40
Traffic type Constant Bit Rate (CBR)
Topology 3000 m ∗ 3000 m
Transport protocol UDP
Packet size 512 Bytes
Routing protocol AODV
Channel type SCH
Number of road lanes 2
Radio propagation model Two ray ground
MAC protocol IEEE 802.11p
Max speed 15 m/s
Mobility model Manhattan mobility model
Hence, claims itself as an intermediate node in the path from source to destina-
tion node. For generating the dataset, we simulate the network multiple times
with a different pair of nodes as attacking nodes. We label the transmission
occurring between the two attacking nodes as abnormal and the rest as normal.
Parameter Value
Number of features 11
Classes 2 (0-Normal, 1-Abnormal)
Number of instances 11444
Class label 0 10000
Class label 1 1444
Training size 8010
Testing size 3434
Missing values NA
After extracting the features from flow monitor in the previous stage, we need to
assure that each class has balanced class distribution. In our prepared dataset, we
have more than 98% instances labeled normal. We use downsampling technique
658 P. K. Singh et al.
to overcome this issue which randomly eliminates majority class instances from
the dataset. Table 2 shows the characteristics of the final dataset to be used for
learning the attack.
From the feature set, some features may not be correlated with the class
labels. The feature selection algorithm [9] selects a subset of features which gives
maximum accuracy. This method helps in finding the best subset of features
which directly affects the behavior of the system. The straightforward approach
in finding the best subset of features is to apply learning model for every possible
subset of features. The subset on which the model gives the maximum accuracy
is the required set of features. The straightforward approach has a very large
search space. There are many other standard feature selection algorithm. In this
work, we use the Sequential Forward Selection (SFS) algorithm and Sequential
Backward Selection (SBS) algorithm. These two algorithms are greedy in nature.
Figure 4a and 4b show the variation of accuracy with a number of features.
From the experiment, the feature selection algorithm returns the set of fea-
tures, which gives maximum accuracy are Source and destination IP Address,
Transmitted and Received Bytes, Dropped Bytes, F irstRxBytesT ime −
F irstT xBytesT ime and Throughput.
5 Results
In this section, we discuss the results that we achieved after the experiment.
We split the dataset into training and testing data. The training data consists
of 70% of total instances, and testing data contains 30% of total instances. For
KNN, the value of k, the number of nearest neighbors is 4. The value of k is set
to 4 based on accuracy that we got in experiment. The accuracy of the model is
the number of correctly classified test instances out of total test instances. Both
the SVM and k-NN performed well and detection the attack with accuracy of
99%.
We compare the performance of the learning models used to learn the attack
on the basis of accuracy and the different alarm rates. The four different alarm
types are True Positive (TP), True Negative (TN), False Positive (FP) and False
Negative (FN). The different alarm rates for wormhole attack by two models are
shown in Table 3. From learning models, we expect a high accuracy and a low
false alarm rate. From the results, we observe that both the models perform well
for detecting wormhole attack with the high accuracy and a low false alarm rate.
6 Conclusion
In this paper, a machine learning model-based approach is proposed to detect
wormhole attack in VANETs. We demonstrated with our methodology that how
effective ML is in detecting wormhole attack in the network. To get the best set
of features, we use feature selection algorithm. The method proposed to use a
self-created dataset to learn the attacks. The described model detects the attack
with high accuracy and low false alarm rate. ML has great potential and can be
utilized in detecting attacks at different layers of the VANET protocol stack.
660 P. K. Singh et al.
References
1. Albouq, S.S., Fredericks, E.M.: Detection and avoidance of wormhole attacks in
connected vehicles. In: Proceedings of the 6th ACM Symposium on Development
and Analysis of Intelligent Vehicular Networks and Applications, pp. 107–116.
ACM (2017)
2. Ali Alheeti, K.M., Gruebler, A., McDonald-Maier, K.: Intelligent intrusion detec-
tion of grey hole and rushing attacks in self-driving vehicular networks. Computers
5(3), 16 (2016)
3. Altman, N.S.: An introduction to kernel and nearest-neighbor nonparametric
regression. Am. Stat. 46(3), 175–185 (1992)
4. Argyroudis, P.G., O’mahony, D.: Secure routing for mobile ad hoc networks. IEEE
Commun. Surv. Tutor. 7(1–4), 2–21 (2005)
5. Bakhouya, M., Gaber, J., Lorenz, P.: An adaptive approach for information dis-
semination in vehicular ad hoc networks. J. Netw. Comput. Appl. 34(6), 1971–1978
(2011)
6. Behrisch, M., Bieker, L., Erdmann, J., Krajzewicz, D.: SUMO–simulation of urban
mobility. In: The Third International Conference on Advances in System Simula-
tion (SIMUL 2011), Barcelona, Spain, vol. 42 (2011)
7. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273–297
(1995)
8. Grover, J., Prajapati, N.K., Laxmi, V., Gaur, M.S.: Machine learning approach
for multiple misbehavior detection in VANET. In: International Conference on
Advances in Computing and Communications, pp. 644–653. Springer (2011)
9. Guyon, I., Elisseeff, A.: An introduction to variable and feature selection. J. Mach.
Learn. Res. 3(Mar), 1157–1182 (2003)
10. Henderson, T.R., Lacage, M., Riley, G.F., Dowell, C., Kopena, J.: Network simu-
lations with the ns-3 simulator. SIGCOMM Demonstr. 14(14), 527 (2008)
11. Hu, Y.C., Perrig, A., Johnson, D.B.: Packet leashes: a defense against wormhole
attacks in wireless networks. In: Twenty-Second Annual Joint Conference of the
IEEE Computer and Communications, INFOCOM 2003. IEEE Societies, vol. 3,
pp. 1976–1986 IEEE (2003)
12. Hu, Y.C., Perrig, A., Johnson, D.B.: Wormhole attacks in wireless networks. IEEE
J. Sel. Areas Commun. 24(2), 370–380 (2006)
13. IEEE Std.: IEEE standard for information technology– local and metropolitan area
networks – specific requirements– Part 11: wireless LAN medium access control
(MAC) and physical layer (PHY) specifications amendment 6: wireless access in
vehicular environments, pp. 1–51, July 2010
14. Johnson, D.B., Maltz, D.A., Broch, J., et al.: DSR: the dynamic source routing
protocol for multi-hop wireless ad hoc networks. Ad Hoc Netw. 5, 139–172 (2001)
15. Kang, M.J., Kang, J.W.: Intrusion detection system using deep neural network for
in-vehicle network security. PloS One 11(6), e0155781 (2016)
16. Khalil, I., Bagchi, S., Shroff, N.B.: Liteworp: detection and isolation of the worm-
hole attack in static multihop wireless networks. Comput. Netw. 51(13), 3750–3772
(2007)
17. Loukas, G., Vuong, T., Heartfield, R., Sakellari, G., Yoon, Y., Gan, D.: Cloud-
based cyber-physical intrusion detection for vehicles using Deep Learning. IEEE
Access 6, 3491–3508 (2018)
18. Perkins, C., Belding-Royer, E., Das, S.: Ad hoc on-demand distance vector (AODV)
routing. Technical report (2003)
Machine Learning Based Approach to Detect Wormhole Attack in VANETs 661
19. Singh, P.K., Dash, M.K., Mittal, P., Nandi, S.K., Nandi, S.: Misbehavior detec-
tion in C-ITS using deep learning approach. In: 18th International Conference on
Intelligent Systems Design and Applications (ISDA). Springer (2018)
20. Sanzgiri, K., Dahill, B., Levine, B.N., Shields, C., Belding-Royer, E.M.: A secure
routing protocol for ad hoc networks. In: 10th IEEE International Conference on
Network Protocols, pp. 78–87. IEEE (2002)
21. Singh, P.K., Sharma, S., Nandi, S.K., Nandi, S.: Multipath TCP for V2I commu-
nication in SDN controlled small cell deployment of smart city. Veh. Commun. 15,
1–15 (2018)
22. Taylor, A., Leblanc, S., Japkowicz, N.: Anomaly detection in automobile control
network data with long short-term memory networks. In: 2016 IEEE International
Conference on Data Science and Advanced Analytics (DSAA), pp. 130–139. IEEE
(2016)
Secure Peer-to-Peer Communication
Based on Blockchain
1 Introduction
Public Key Infrastructure (PKI) is a significant component to resolve authen-
tication in networks and provides guarantees to trust a certificate signed by a
certification authority (CA). A new concept named web of trust for his Pretty
Good Privacy (PGP) encryption program uses PKI to provide confidentiality
with encryption, authentication via the signature and web of trust via identity
validation from peers. The certificates authenticate the public keys and allow
you to perform cryptographic operations, such as encryption and digital signing.
As authentication and identity validation is centralized in the PKI, that creates
single points of failure.
The blockchain technology was designed to make the transaction more reliable.
Blockchain technology is a potential solution to achieve data integrity, relying on
cryptography to provide tamper-resistance. Blockchain can be utilized in such
cases to achieve the secure communication and integrity of data. The blockchain is
decentralized, and no centralized authority can approve transactions. All network
participants must reach a consensus to validate transactions in a secure way, and
previous records cannot be changed. A very high cost must be spent if someone
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 662–672, 2019.
https://doi.org/10.1007/978-3-030-15035-8_64
Secure Peer-to-Peer Communication Based on Blockchain 663
wants to alter previous records. External attackers would have to gain access to
every computer in the network that hosts the blockchain database at the same
time to manipulate it, which is as practically impossible [1].
The cryptocurrency Bitcoin was created by an unknown person using the
alias Satoshi Nakamoto in 2008. Bitcoin generates and broadcast transactions
to the Blockchain. Once validated, the transactions are propagated through the
network and added into a block. Once a block is full, the block is appended
to the Blockchain by performing a mining process. Miners try to solve a hard
cryptographic puzzle named Proof of Work (PoW), and the node that solves the
puzzle first adds the new block to the Blockchain. Bitcoin is based on decen-
tralized trust. The trust is achieved as an emergent property from the interac-
tions of different participants in the bitcoin system [2]. The data stored on the
blockchain cannot be hacked, modified or deleted. The immutability of data in
this blockchain is strongest when the chain is long [3].
1.1 Motivations
Blockchain has been of great interest to engineers and investors because of its
immense commercial potential and its use in applications as diverse as a cryp-
tocurrency.
Several layering schemes of new features on bitcoin are proposed. A proposal
to create a decentralized domain name service over Bitcoin appeared as BitDNS,
eventually evolved into Namecoin, the first altcoin [4]. Bitcoin has the most sub-
stantial amount of computational power securing the blockchain data. However,
it’s hard to introduce new functionality to bitcoin because that requires consen-
sus breaking changes. Transactions in bitcoin did not include a standard method
for carrying a data payload, and the blockchain grows at an exponential rate [5]
that extends pressures upward on storage space and importantly network band-
width. Due to the security concerns, only a handful of functions are permitted
in standard transactions [6].
Peer to peer (P2P) systems help overcome many problems that go beyond
traditional client-server approaches, but these features also introduce new issues
such as how to establish the trust relationship within P2P networks [7].
The Ethereum blockchain is a platform for decentralized applications called
smart contracts. Because Ethereum addresses are unique identifiers whose own-
ership does not change, their activity can be tracked and analyzed [8]. A smart
contract is an executable code that runs on top of the blockchain, allows run-
ning an agreement between two parties automatically, without one of the parties
being able to obstruct its execution [9].
At the time of the digital economy, data is brought to transit more and
more between companies, from a client to a supplier, moving from one cloud to
another with the advent of virtualization and containers. Our work describes the
potential for applying blockchain to assure the traceability of data, certificate
individual and secure messaging based on the blockchain technology.
664 K. Khacef and G. Pujolle
In this contribution, we are interested in the security of the blockchain for mes-
saging, which can be an email, a website or some other form of message. We
propose a protocol that performs encrypted messaging on the public blockchain.
The goal of our contribution is to achieve the three functionalities achieved by
PKI using the blockchain as the database to store public keys, digital signature,
and peer information, allowing each entity of the network to validate information
about every other node in the network.
The main contribution of this paper is: Sect. 2 a presentation of the benefits
of blockchain in cybersecurity, Sect. 3 briefly description of the current state of
the art for Blockchain based PKI, and in Sect. 4 we present our proposed solution
with a smart contract to use the legitimate secure certificate.
2 Blockchain
the blockchain nodes is used. However, raise their questions, Who has the
authority to grant permission? A permission blockchain may make its owners
feel more secure, giving the database rigorous security and privacy capabilities
but can see as violating the idea of blockchain because only some participants
have more control, which means they can make changes whether or not other
network participants agree.
3 Related Work
In this section, we present the works made use Blockchain for secure transactions
between users. Since BitDNS [4], numerous approaches are proposed.
Namecoin it is the first system to build a decentralized naming system using
blockchain, the first altcoin from bitcoin with its blockchain [14]. Satoshi believed
that BitDNS should use its independent blockchain and event offered the first
proposal for merged mining as a way to secure blockchain. Namecoin was also
the first solution to zooko’s triangle producing a naming system that is simul-
taneously secure, decentralized, and human-meaningful. Namecoin suffers from
51% attacks [17], because of its insufficient computing power.
Blockstack uses the existing internet transport layer (TCP or UDP)
and underlying communication protocols and focuses on removing points of
Secure Peer-to-Peer Communication Based on Blockchain 667
4 Contribution
4.1 The Proposed Model
The primary goal of our approach is to secure communications between enti-
ties of the network. The proposed model is to use the blockchain to validate
the user’s identity and to ensure trust between users for exchanging messages
with a high level of security. Each user must communicate only with the user’s
identity validated by the smart contract, and consider every other interaction
as malicious. Each user how want uses the system to communicate with others
must register their identity and their public key an store it on blockchain. To
achieve such a system, we can use Ethereum public blockchain that implements
smart contracts. The Ethereum network has up to 200,000 developers currently
working on the protocol and derived projects, ensuring that the protocol remains
Table 1. Notations
Notation Definition
Alice, Bob Entities
userA
Kpub Public key of Alice
userB
Kpub Public key of Bob
userA
Kpri Private key of Alice
userB
Kpri Private key of Bob
A
ID Id of Alice
IDB Id of Bob
userA
Spri Signature with private key of Alice
userA
Epub Encryption with public key of Alice
t Validity of the Public Key
T Time stamps
668 K. Khacef and G. Pujolle
cutting edge, it has a level of security. The most frequently used notations in
our scheme are mentioned in Table 1.
Once the code conditions are validated in smart contract, it can run as sched-
uled. This execution is carried out in practice by sending a specific transaction
to the blockchain that causes the execution of the smart contract (Fig. 1).
Fig. 2. ID generation
4.2 Evaluation
4.2.1 Advantages
Our approach removes central authorities (CA) and uses the blockchain pub-
lic as a distributed ledger of identity and their associated public keys. We use
Blockchain to store public keys, digital signature, and peer information.
Once published, the smart contract code works precisely as programmed.
This is one of the main advantages of the platform, the code always interacts
as promised it cannot be falsified, and it never has a downtime. The system is
trust, transparent and traceable.
4.2.2 Limitations
In blockchain, the smart contract is executed sequentially that affect the per-
formance of blockchain negatively. With the growing number of smart contract,
the blockchain will not be able to scale.
In practice, it is impossible to modify an existing contract in the blockchain
after it has been registered in it. The design phase of these contracts will, there-
fore, require special attention to avoid any future disappointment.
4.3 Implementation
The smart contract plays a vital role in executing and performing the agree-
ment among various users in the system. For implementing the proposed sys-
tem, a smart contract can be created by developing the codes that run on the
Secure Peer-to-Peer Communication Based on Blockchain 671
blockchain to execute the agreement signed by the nodes to validate the certifi-
cates of individuals. Smart contract promise low transaction fees compared to
traditional systems that require a trusted third party to execute an agreement.
The contract’s state is stored on the blockchain, and it is updated each time the
contract is invoked.
Before sending/receiving the message, the smart contract first checks the
identity and its associated public key, previously stored in the blockchain. Then,
it checks the validity of the timestamp. Finally, the smart contract verifies the
signature.
A smart contract can be developed with Solidity.
5 Conclusion
Our contribution benefits from an entirely decentralized architecture offering
inherent fault tolerance, redundancy, and transparency. We have firstly proposed
an approach that secure communication, it benefits of security properties of
blockchain public. It shows how to use to the immutability of the blockchain to
provide a solution to high problematics in the field of centralized PKI.
We plan to implement our proposal to base our results empirically and
demonstrate its viability.
Our next proposal is to set up an architecture with smart contracts to val-
idate, store and revoke the certificate on a public blockchain. The certificate
of the individual will contain, his address and public key, the address of smart
contract that issued it, and stored it in Off-chain.
References
1. Schuermann, F.: Bitcoin and beyond: a technical survey on decentralized digital
currencies. IEEE Commun. Surv. Tutor. 18, 2084–2092 (2016)
2. Bano, S., Sonnino, A., Bassam, M., Azouvi, S., McCorry, P., Meiklejohn, S.,
Danezis, G.: SoK: consensus in the age of blockchains, arxiv (2016)
3. Cachin, C., Vukoli, M.: Blockchain Consensus Protocols in the Wild (2016)
4. Loibi, A.: Namecoin, Seminar innovative Internettechnologien und mobilkommu-
nikation SS2014. IEEE (2014)
5. Hung, T., Svetinovic, D.: Data analysis of digital currency network: namecoin case
study. In: International Conference on Engineering of Complex Computer Systems
(2016)
6. Deepak, K., Shetty, S., Liang, X., Kamhoua, C., Njilla, L.: Consensus Protocols for
Blockchain-based Data Provenance: Challenges and Opportunities. IEEE (2017)
7. Mudliar, K., Parekh, H., Bhavathankar, P.: A comprehensive integration of national
identity with blockchain technology. In: 2018 International Conference on Com-
munication, Information and Computing Technology (ICCICT), 2–3 February,
Mumbai, India (2018)
8. Aung, Y.N., Tantidham, T.: Review of Ethereum: smart home case study. In: 2017
2nd International Conference on Information Technology (INCIT) (2017)
9. Ward, M.: Untangling Blockchain: A Data Processing View of Blockchain Systems,
Elsevier Information Security Technical report, pp. 89–92 (2006)
672 K. Khacef and G. Pujolle
10. Tewari, H., Nuallain, E.: Netcoin: A Traceable P2P Electronic Cash System. IEEE
(2016)
11. Guy, Z., Oz, N., Alex, P.: Decentralized Privacy: Using Blockchain to Protect
Personal Data (2018)
12. Conti, M., Kumar, S., Lal, C., Ruj, S.: A Survey on Security and Privacy Issues of
Bitcoin. IEEE (2017)
13. Tosh, D.K., Shetty, S., Liang, X., Kamhoua, C., Njilla, L.: Consensus Protocols for
Blockchain based Data Provenance: Challenges and Opportunities. IEEE (2017)
14. Kalodner, H., Carlsten, M., Ellenbogen, P., Bonneau, J., Narayanan, A.: An empir-
ical study of Namecoin and lessons for decentralized namespace design. Blockchain
(2014)
15. Sankar, L.S., Sindhu, M., Sethumadhavan, M.: Survey of consensus protocols on
blockchain applications. In: International Conference on Advanced Computing and
Communication Systems (ICACCS 2017)
16. Mingxia, D., Xiaofeng, M., Zhe, Z., Xiangwei, W., Qijin, C.: A Review on Consen-
sus Algorithm of Blockchain. IEEE (2017)
17. Dai, F., Shi, Y., Meng, N., Wei, L., Ye, Z.: From bitcoin to cybersecurity: a com-
parative study of blockchain application and security issues/ICSAI, pp. 975–979
(2017)
18. Ali, M., Nelson, J., Shea, R., Freedman, M.J.: Blockstack: a global naming and
storage system secured by blockchain. In: USENIX Annual Technical Conference
(2016)
19. Fronknecht, C., Velicannu, G., Sophia, Y.: CertCoin: A Namecoin Based Decen-
tralized Authentication System (2014)
20. Khovayko, O., Shumilov, E.: EMCSSL Decentralized identity management, pass-
wordless logins, and client SSL certificates using Emercoin NVS (2016). http://
emercoin.com
21. Andrea, C.: ECDH and ECDSA (2015). https://andrea.corbellini.name/2015/05/
30/elliptic-curve-cryptography-ecdh-and-ecdsa
Developing the Analysis Tool
of Cyber-Attacks by Using CTI
and Attributes of Organization
1 Introduction
1.1 Background
Cyber-attacks range over various area such as Web services, IoT and Cryptocur-
rency. Moreover, there are a lot of ways to attack. Recently, a time period from
beginning of a cyber-attack to achieving its goal, has been shortened. Hence,
preliminary measures and countermeasures that focus on prevention of dam-
age diffusion are necessary. To protect an organization such as infrastructure
facilities and government from cyber-attacks, we have to collect the information
about cyber-attacks in the world such as what kind of attacks are happened,
how it is performed and what kind of information are targeted. If we get a valid
information quickly for organization, we can make the countermeasures effective.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 673–682, 2019.
https://doi.org/10.1007/978-3-030-15035-8_65
674 Y. Kambara et al.
Today, a lot of cyber-attacks are happened in the world and the amount of
CTI data increases in proportion. CTI data enables us to detect cyber-attacks
and prevent diffusion of damage. However, it is difficult for an organization
to find the useful CTI data from a lot of CTI, because effective CTI data for
the organization is very few. We have to extract the CTI data related to the
organization from massive CTI to utilize it. Also, to determine a countermeasure,
we have to analyze cyber-attacks. There is a possibility that a cyber-attack which
seems to be new at first glance is similar to existing attacks. In that case, the
past measures may be applied to the new attack. Hence, We need a mechanism
that reveals whether the past measures can be applied to the new attack by
judging a similarity between the past CTI data and current CTI data.
Then, we paid attention to attributes of the organization. It becomes possible
to extract the related CTI from a large amount of CTI by using this information.
We developed an analysis tool which can pick up and visualize a useful CTI
information and the information of organization as a graph. This tool enables us
to analyze the relevancy between CTI and the organization and reconciliation.
In this paper, we explain the method of efficiently analysis of CTI.
2 Data Set
2.1 Cyber Threat Intelligence
Observables:
The information about evidence of cyber-attack such as adversary’s server
information (IP addresses, Domain names and URLs) and a malware infor-
mation (hash values, sizes and names). Observables are used synonymously
with IOC.
TTPs:
Information about an adversary including methods of a cyber-attack and
targeted systems.
Threat Actors:
Adversary’s motivations, intentions and information about people/
organization that were involved in a cyber-attack.
Courses of Action:
The information about countermeasures to prevent from threat. There are a
lot of information such as circumstances of countermeasures, influence of the
organization, and effectiveness of countermeasures.
There are many methods to get STIX data, for example, TAXII protocol,
which is used exclusively for STIX. The protocol was standardized in 2014.
According to the official documents of STIX, CTI data written in STIX can
describe information including the adversary’s attack methods, the motivation,
the purpose, the name of threat group and how to deal with the cyber-attack.
We investigated what types of information we can acquire by downloading CTI
data from OTX, HAT and one of the security vendors. Table 1 shows the list of
the data we acquired.
Most of the CTI data we actually acquired is an Observables type. Also,
there exists a little information that is available only for specific service. For
example, the YARA rule used for malware detection can be acquired from OTX
and the port number used in illegal access can be acquired from HAT. In addition
to the Observables, OTX is providing the industry and geography information
of the targeted organization that received damage (TTPs) and threat group
data (Threat Actors). The data of methods and purposes of cyber-attacks were
676 Y. Kambara et al.
acquired only from the security vendor’s CTI, but there are no data of the coping
method of cyber-attacks (Courses of Action).
Ideally, an organization on defense side has to check all of the provided CTI data
whether their corresponding threats are appropriately dealt with or not. How-
ever, a large amount of CTI data exists in the Internet and CTI data valuable to
a given organization is rare. Therefore, it takes a long time for an organization
to find CTI data related to directions or operations of its business. To suppress
the damage of cyber-attacks, we have to extract CTI data related to the orga-
nization from a large amount of data and handle quickly CTI data with high
priority. Then, we use the information of organization to reveal a relationship
between CTI data and the organization.
Threat levels of cyber-attacks are different for each organization. For example,
let’s assume there are two adversaries shown in Table 2, and three organizations
shown in Table 3.
Developing the Analysis Tool of Cyber-Attacks 677
For Company A, we can see CTI1 has a high relevance, because the company
has a medical information and located in Japan, and organizations with such
attributes are targeted by CTI1. On the other hand, company A is not targeted
by the adversary of CTI2. Hence, company A has to cope with the CTI1 with
high priority. Company C has to cope with the adversary of CTI2 likewise. In
this way, similarities between adversary’s purpose and organization become an
indicator to attach a priority to several CTI data. We suppose that organization
information has following items:
We sort out these pieces of information and connect them with CTI data.
Table 4 shows corresponding items of CTI data and the information of organiza-
tion. In modern days, these pieces of information are available in the Internet,
which are used by adversaries to attack organizations.
In this research, we developed an analysis tool of CTI which has the following
functions:
It is difficult for the organization to find the useful CTI from a large amount of
CTI. We extract the CTI data which is related to the organization by using the
attribute of it. Then, we calculate the priority of CTI data. Calculating method
is the following. We use “target geography” and “target information” from CTI
and the information of organization. We define Iorg as the set of the attribute
of the organization, Icti as the set of the information which are targeted in the
CTI, Gorg as the set of the geography which locates the organization and Gcti
as the set of targeted geography in the CTI. Then, the priority is defined below.
In Fig. 1, there are 62 CTI data (blue frame in left). After selection by the
information of organization, the number of CTI is reduced to 4 (blue frame in
right). These remained CTI are related to the organization. To do this, security
operator of the organization can select the CTI which has high priority for them,
and proceed handling of the CTI.
Visualizing the attribute of the organization and CTI data enables us to under-
stand the relationship between these two. Figure 2 is an image of the relationship
between attribute of the organization and CTI data. CTI data is in the green
frame. The attribute of the organization is in the red frame. The organization
has to cope with Observable (Fig. 2, right), for example, access blocking. Figure 2
shows the cyber-attack that targets 5 countries (bottom left) and various infor-
mation (orange node) such as Financial information. The organization (in the
red frame) based in 3 locations (Japan, China and United States) and has the
one of the information that targeted by the adversary. Therefore, we can know
that the adversary targets this organization and the organization has to take
a measure. By combining CTI data with the information of organization, we
can understand the relationship between the organization and CTI data at a
glance. In other words, we will know the threat which is the most dangerous in
680 Y. Kambara et al.
Fig. 3. CTI witch has high (low) priority for the organization
a short time. Figure 3 shows high priority CTI data and low priority CTI data.
The attribute of the organization and CTI data on the right side has 4 com-
mon nodes. On the other hand, CTI data on the left side has 1 common node.
This indicates that the strength of the relationship between the organization
and right CTI is higher than the other. Hence, we have to cope with CTI on
the right side. Moreover, the red line shows the relevancy among CTI and we
can judge the similarity. If there are two or more CTI, it is possible to apply
the same countermeasure to these CTI. Hence, to judge the similarity enables
to determine countermeasure rapidly.
Developing the Analysis Tool of Cyber-Attacks 681
• Putting different hash values which do the same behavior together. This helps
to recognize the malware family.
• Putting different CTI which caused by the same adversary. This helps to
grasp tendency of cyber-attacks together.
Analyzing the fuzzy relation between plural CTI is possible by storing the result
of analysis by human. Also, this helps us to take sophisticated countermeasures.
682 Y. Kambara et al.
5 Conclusion
Organization information enables us to cope with cyber-attacks in the order of
descending priorities. We can also extract an useful data which is related to an
organization from a large amount of CTI data. It is easy to check countermeasure
circumstance of CTI data and also easy to analyze a similarity among them by
visualizing it and the information of organization. Future works include verifying
usefulness of our tool and improving it to make a countermeasures more effective.
References
1. AlienVault, Open Threat Exchange (OTX). https://www.alienvault.com/open-
threat-exchange
2. Gartner, Reviews for Security Threat Intelligence Products and Services. https://
www.gartner.com/reviews/market/security-threat-intelligence-services
3. HAIL A TAXII. http://hailataxii.com/
4. OASIS, Introduction to STIX. https://oasis-open.github.io/cti-documentation/
stix/intro
5. OASIS, Introduction to TAXII. https://oasis-open.github.io/cti-documentation/
taxii/intro
6. Masuoka, R., Satomi, T., Yamada, K.: Strategies for semantics-rich machine-
understandable CTI - human - machine cybersecurity collaboration. In: Borderless
Cyber Conference and FIRST Technical Symposium (2017)
7. Katayama, Y., et al.: Use of cyber threat intelligence for organizations. In: Japanese,
Symposium on Cryptography and Information Security (SCIS) (2017)
8. Kambara, Y., et al.: How to set the expiration date of cyber threat intelligence. In:
Japanese, Symposium on Cryptography and Information Security (SCIS) (2017)
Proposal of Ad-Hoc Secure Device
Pairing Method Using Similarity Between
Marker Movement and Acceleration
1 Introduction
Recently, along with the advance of wireless technologies, such as Wi-Fi, Blue-
tooth, and near-field communication (NFC), devices such as mobile device and
IoT devices often use these wireless technologies to communicate with each other.
These technologies have privacy vulnerabilities, such as eavesdropping and man-
in-the-middle attack. Hence, it is necessary to establish authenticity between
devices before wireless communication begin. We define secure device pairing as
a process for establish secure wireless communication.
An example of pairing is the wireless distribution of meeting materials from
a PC to mobile devices with the small table (e.g. 3 m × 3 m table) at a meet-
ing. In this pairing, it is necessary to generate a pairing key “on-the-spot”.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 683–692, 2019.
https://doi.org/10.1007/978-3-030-15035-8_66
684 M. Nagatomo et al.
The problem in this case is spoofing by a third party outside the room, who can
obtain the meeting materials if secure pairing is not performed between devices
in the room. When Wi-Fi or Bluetooth is used for it, the wireless communica-
tion may be eavesdropped. Therefore, our research goal is to propose the secure
device pairing method, in which is easy to use for users, and it is possible to
perform the pairing in short-range.
Currently, many researches evaluate pairing methods using received signal
strength (RSS) [1,2]. However, it is not possible to perform stable pairing to a
single location because RSS changes significantly due to environmental factors,
such as time and objects.
There is the method which a device having function of emitting visible light
conveys information to the device equipped with a camera [3]. However, the
pairing distance of these methods are short. An alternative is proximity pairing,
in which an infrared camera recognize user’s hand which holds a hand-held device
[4,5]. However, this method requires a special infrared camera. Thus, method
using a regular camera are proposed [6,7]. These methods are tracking methods
of objects with accelerometer. However, these methods is not secure because
camera cannot recognize movements of the objects in detail.
Therefore, we proposed secure device pairing method using similarity between
displacement data of a marker displayed on devices and acceleration data
obtained from devices [8], and performed experiments to confirm similarity and
whether impersonator can pair with the PC. However, we founded that similarity
was not unstable.
In this paper, we propose extended method which marker changes at an
interval. These sequence of markers displayed on the device is used to distin-
guish a device from some devices by a camera in addition to calculation of the
similarity. In addition, we perform three experiments to confirm the similarity
displacement data and acceleration data, and whether an eavesdropper outside
the camera range can perform pairing and perform three device pairing together.
In Sect. 2, we review related works. Section 3 gives a system model and pair-
ing procedure of proposed method. In Sect. 4, we evaluate the accuracy of the
proposed method. Finally, Sect. 5 concludes this research work.
2 Related Work
Amigo [1] is a RSS-based method for pairing a PC and a device. This method
perform identification of a device using machine learning for the features of mean
absolute difference of RSS, mean exponential of difference of RSS, and euclidean
difference of RSS vectors. This method can detect an eavesdropper when the
distance between the attacker and the legitimate devices is 3 m. In addition,
device pairing method in a room using RSS from some APs was proposed [2].
Machine learning is used for the features of mean absolute difference of RSS from
beacon frames in 2.4-GHz band and 5-GHz band obtained from shared APs and
all APs of two devices. This method judge whether devices is in the same room
of the 10 to 15 m2 . An accuracy was higher than 96%. However, RSS varies easily
Proposal of Ad-Hoc Secure Device Pairing Method 685
Figure 1 shows the system model of our proposed method. The components of
this model are as follows.
where ṽiw , ũi (w ∈ {x, y, z}, u ∈ {x, y}, i ∈ {1, ..., m}) are calculated as follows:
m
m
i
ṽw = viw / (vkw )2 , ũi = ui / u2
k, (7)
k=1 k=1
Benefits: This method has longer pairing distance than QR code method
because a marker of this method can be more simple than that of QR code
method. Hence, one-to-many device pairing is possible by reading markers on
the displays of several devices.
In our proposed method, server S calculates the sequences of markers
obtained from camera and device in step 8. S calculates the similarity ms as
follows:
β
number that mα
i = mi in α̃ and β̃
ms = (8)
number of set α̃ or β̃(m)
equation.
m
(ṽiw − ṽ w )(w̃i − w̃)
rw = m i=1 m , (9)
( i=1 (ṽiw − ṽ w )2 )( i=1 (w̃i − w̃)2 )
m m
where w ∈ {x, y}, ṽ w = k=1 ṽkw , and w̃ = k=1 w̃k . rw takes a value from −1
to 1. It can be judged that there is a positive correlation when rw is positive.
When rw is negative, there is a negative correlation.
1. The student holds the device 1.0 m away from the camera and launches the
device application.
2. The student holds the device for the laptop camera. The student moves the
device along the shape of a circle or ∞ symbol for about 5 s.
Proposal of Ad-Hoc Secure Device Pairing Method 689
(a) Velocity data converted from displacement (b) Velocity data converted from acceleration
data data
There is a possibility that the third party who is outside the range of the camera
imitates the movement of the legitimate user. If the similarity has variability,
690 M. Nagatomo et al.
Device, motion Nexus 5X, circle Nexus 5X, ∞ Nexus 7, circle Nexus 7, ∞ Average
Average similarity 0.717 0.688 0.855 0.854 0.778
Standard deviation 0.248 0.217 0.213 0.199 0.232
Device, motion Nexus 5X, circle Nexus 5X, ∞ Nexus 7, circle Nexus 7, ∞ Average
Average similarity 0.571 0.575 0.586 0.619 0.588
Standard deviation 0.265 0.223 0.292 0.197 0.197
that of the impersonating device can be higher than that of legitimate device.
Therefore, we perform the confirmation experiment for considering it. Subjects
are 12 university students. We performed this experiment as following procedure.
1. Legitimate user holds the device 1.0 m away from the PC’s camera. Both start
the device’s application.
2. The user holds the device for the camera. After that, user moves the device
along the shape of circle or ∞ symbol for 5 s. Simultaneously, impersonator
imitates the movement of the user’s device.
3. The PC calculates the similarity between the displacement data obtained in
step 2 and the acceleration data which is transmitted from both user’s device
and impersonator’s device in step 2.
4. We repeat 1–3 five times.
5 Conclusion
References
1. Varshavsky, A., Scannell, A., LaMarca, A., de Lara, E.: Amigo: proximity-based
authentication of mobile devices. In: Proceedings of the 9th International Confer-
ence on Ubiquitous Computing (UbiComp 2007), pp. 253–270 (2007)
2. Agata, Y., Hong, J., Ohtsuki, T.: Room-level proximity detection based on RSS of
dual-band Wi-Fi signals. In: Proceedings of 2016 IEEE International Conference
on Communications (ICC) (2016)
3. Saxena, N., Ekberg, J.E., Kostianien, K.: Secure device pairing based on a visual
channel: design and usability study. IEEE Trans. Inf. Forensics Secur. 6(1), 28–38
(2010)
4. Rofouei, M., Wilson, A.D., Brush, A.J.B., Tansley, S.: Your phone or mine? Fus-
ing body, touch and device sensing for multi-user device-display interaction. In:
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems,
pp. 149–158 (2012). Camera View, Intelligent Robots and Systems, pp. 3872–3877
(2008)
5. Wilson, A.D., Benko, H.: CrossMotion: fusing device and image motion for user
identification, tracking and device association. In: Proceedings of the 16th Inter-
national Conference on Multimodal Interaction, pp. 216–223 (2014)
6. Maruhashi, N., Terada, T., Tsukamoto, M.: A method for identification of moving
objects integrative use of a camera and accelerometers. In: Proceedings of the 27th
Annual ACM Symposium on Applied Computing, pp. 1–6 (2012)
7. Osamu, S., Kagami, S., Hashimoto, K.: Identifying a moving object with an
accelerometer in a camera view. In: Proceedings of the 16th International Con-
ference on Multimodal Interaction, pp. 216–223 (2014)
8. Nagatomo, M., Aburada, K., Okazaki, N., Park, M.: Proposal and evaluation of
secure device pairing method with camera and accelerometer. In: Proceedings of
2018 Sixth International Symposium on Computing and Networking Workshops
(CANDARW), pp. 310–315 (2018)
9. Diffie, W., Hellman, M.E.: New directions in cryptography. IEEE Trans. Inf. The-
ory 22(6), 644–654 (1976)
10. Aruco. https://docs.opencv.org/3.2.0/d5/dae/tutorial aruco detection.html. Acc-
essed 28 Jan 2019
Introduction of Fingerspelling
for Realizing a User Authentication
Method Using s-EMG
1 Introduction
This paper presents an introduction of fingerspelling as candidates of gestures
used in the user authentication method for mobile devices by using surface elec-
tromyogram (s-EMG) signals, not screen touching.
An authentication method that prevents shoulder surfing, which is the direct
observation of a users personal information such as passwords, comes to be
important. At the present time, mobile devices such as tablet type PCs and
smartphones have widely penetrated into our daily lives. So, authentication oper-
ations on mobile devices are performed in many public places and we have to
ensure that no one can view our passwords. However, it is easy for people who
stand near such a mobile device user to see login operations and obtain the users
authentication information. And also, it is not easy to hide mobile devices from
attackers during login operations because users have to see the touch screens of
their mobile devices, which do not have keyboards, to input authentication infor-
mation. On using a touchscreen, users of a mobile device input their authentica-
tion information through simple or multi-touch gestures. These gestures include,
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 693–701, 2019.
https://doi.org/10.1007/978-3-030-15035-8_67
694 H. Yamaba et al.
for example, designating his/her passcode from displayed numbers, selecting reg-
istered pictures or icons from a set of pictures, or tracing a registered one-stroke
sketch on the screen. The user has to see the touch screen during his/her login
operation; strangers around them also can see the screen.
To prevent this kind of attack, biometrics authentication methods, which use
metrics related to human characteristics, are expected. In this study, we inves-
tigated application of surface electromyogram (s-EMG) signals for user authen-
tication. s-EMG signals, which are detected over the skin surface, are generated
by the electrical activity of muscle fibers during contraction. These s-EMGs
have been used to control various devices, including artificial limbs and electri-
cal wheelchairs. Muscle movement can be differentiated by analyzing the s-EMG
[1]. Feature extraction is carried out through the analysis of the s-EMGs. The
extracted features are used to differentiate the muscle movement, including hand
gestures.
In the previous researches [2–7], we investigate the prospect of realizing an
authentication method using s-EMGs through a series of experiments. First,
several gestures of the wrist were introduced, and the s-EMG signals generated
for each of the motion patterns were measured [2]. We compared the s-EMG
signal patterns generated by each subject with the patterns generated by other
subjects. As a result, it was found that the patterns of each individual subject are
similar but they differ from those of other subjects. Thus, s-EMGs can confirm
ones identification for authenticating passwords on touchscreen devices. Next,
a method that uses a list of gestures as a password was proposed [3,4]. And
also, a series of experiments was carried out to investigate the performance of
the method extracting feature values from s-EMG signals adopted in [5] and the
methods to identify gestures using dynamic data warping (DTW) [6] and support
vector machines (SVM) [7]. The results showed that the methods are promising
to identify gestures using extracted feature values from s-EMG signals.
In this paper, we adopted figerspelling as candidates of gestures used in this
authentication method. To prepare enough number of gestures that are needed
to compose a password, which is a list of gestures, The Japanese Sign Language
syllabary were adopted and examined their characteristics.
user may tap other icons in the shift position on the touchscreen, as indicated
by a shift value, to unlock the device. By keeping the shift value secret, people
around the user cannot know the ture pass-icons, although they can still watch
the tapping operation. The rhythm authentication method [9] relieves the user
from looking at the touchscreen when unlocking the device. In this method, the
user taps the rhythm of his or her favorite music on the touchscreen. The pattern
of tapping is used as the password. In this situation, the users can unlock their
devices while keeping them in their pockets or bags, and the people around them
cannot see the tap operations that contain the authentication information.
Adopting s-EMG signals for authentication of mobile devices has three advan-
tages. First, the user does not have to look at his/her device. Since the user can
make a gesture that is used as a password on a device inside a pocket or in a
bag, it is expected that the authentication information can be concealed. No
one can see what gesture is made. Next, it is expected that if another person
reproduces a sequence of gestures that a user has made, the authentication will
not be successful, because the extracted features from the s-EMG signals are
usually not the same between two people. And then, a user can change the list
of gestures in our method. This is the advantages of our method against other
Introduction of Fingerspelling for a User Authentication Method 697
biometrics based methods such as fingerprints, an iris, and so on. When authen-
tication information, a fingerprint or an iris, come out, the user can’t use them
because he/she can’t change his/her fingerprint or iris. But the user can arrange
his/her gesture list again and use the new gesture list.
used to help the communication using sign languages. For example, proper nouns
such as names of persons are represented by fingerspelling. Sign languages such
as American Sign Language, French Sign Language, British Sign Language, and
so on, have there own manual alphabet.
We adopted the Japanese Sign syllabary (see Fig. 3) as the candidates of
gestures used in our authentication method. They are called yumimoji, which
means “finger letters”. Comparing with manual alphabets representing Latin
alphabet, there are larger numbers of manual kana in yumimoji, 46 letters.
By adopting gestures referring figerspelling, a password that is made up of
gestures corresponds to a string that is made up of letters. It is expected that
this helps users to remember their passwords.
However, some yubimoji may not generate distinct s-EMG signals. To select
yumimoji that are suitable for the use of this study, we examined their perfor-
mances as pass-gesture by measuring their s-EMG signals.
5 Experiments
5.1 Purpose
A series of experiments was carried out to investigate the prospect of the authen-
tication method using yubimoji. Specifically, we investigated whether the mea-
sured s-EMG signals of one experimental subject were similar and whether the
signals of different letters were different from each other.
5.2 Conditions
In these experiments, the set of DL-3100 and DL-141 (S&M Inc.) that was an
electromyograph used in the previous researches also used to measure the s-
EMG of each movement pattern in this study (Fig. 4.) Experimental subjects
Introduction of Fingerspelling for a User Authentication Method 699
made gestures by their left hand. Subjects sit on an arm chair and put his
forearm on the armrest. Two electrode sensors were put on the palm side of the
forearm at the positions of 4 cm and 6 cm toward the elbow from the wrist (The
former is called “m1” and the latter is called “m2” in the following sections).
The measured data were stored and analyzed on a PC.
Five students of University of Miyazaki participated as experimental subjects.
And the 46 yubimoji shown in Fig. 3 were examined. However, since a subject
put his forearm on an armrest, we arranged some yubimoji. When we make a
yubimoji like TE, the palm faces forward. But in these experiment, the palm
faces upward.
Before making each yubimoji, subjects clenched their left fists. s-EMG signals
making yubimoji from a clenched fist were measured. The subjects repeated each
yubimoji ten times and their s-EMG signals were recorded. This measurement
was carried out 3 times and 30 signals were obtained for each subject and for
each yubimoji.
5.3 Results
The number of s-EMG signal data is not so many, but we found that some
yubimoji letters had some caracteristics.
(1) yubimoji that bends one’s wrist forth
The yubimoji of KU, SHI, NI, MI, YO bend one’s wrist forth. Some of their
fingers are extended pointing to the side.
Amplitude of s-EMG signals measured by m2 are large when the yubimoji
is one of this type (Fig. 5(a), (b)).
(2) yubimoji that points downward
The yubimoji of SU, NE, FU, HE also bend one’s wrist forth but their
extended fingers point downward.
Amplitude of s-EMG signals measured by both m1 are as large as (1) and m1
(Fig. 5(d)). Amplitude measured by m2 is also as large as (1) if its middle finger
is extend (SU and NA.) Other (HE and HU) wise they were smaller (Fig. 5(e)).
(3) yubimoji that extend only one finger
The yubimoji of A, I, SE, TA, HI extends only one finger. Also they don’t
bent one’s wrist.
Amplitude of s-EMG signals measured by both m1 and m2 are very slight
(Fig. 5(c)). It seems hard to use these letters as elements of pass-gestures.
(4) yubimoji that extend more than one finger
The yubimoji of U, KE, TE, WA extends more than one finger including
index finger and middle finger. Also they don’t bent one’s wrist.
Amplitude of s-EMG signals measured by m1 are very slight. Amplitude
measured by m2 are larger than those by m1 but smaller than (1) and (2)
(Fig. 5(f)).
From the results of the experiments mentioned above, one’s wrist is bent or
not and some finger is extended or not are expected to be promising clues to
distinction of yubimoji.
6 Conclusion
We investigated a new user authentication method that can prevent shoulder-
surfing attacks in mobile devices. To realize the authentication method using
s-EMG, we examined the characteristics of the yubimoji, the Japanese Sign
Language syllabary, as the candidate of the element of pass-gestures. A series
of experiments was carried out to compare the measured s-EMG signals gener-
ated by making gesutures based on yubimoji. Results of the experiments showed
that which fingers are extended and whether a wrist is bent or not can be used
as characteristics to identify yubimoji from s-EMG signals. We are planning to
examine characteristics of yubimoji using a lot of s-EMG data from many people.
Also, we would like to explore appropriate positions of measuring s-EMG.
References
1. Tamura, H., Okumura, D., Tanno, K.: A study on motion recognition without FFT
from surface-EMG (in Japanese). IEICE Part D J90–D(9), 2652–2655 (2007)
2. Yamaba, H., Nagatomo, S., Aburada, K., et al.: An authentication method for
mobile devices that is independent of tap-operation on a touchscreen. J Robot.
Netw. Artif. Life 1, 60–63 (2015)
3. Yamaba, H., Kurogi, T., Kubota, S., et al.: An attempt to use a gesture control arm-
band for a user authentication system using surface electromyograms. In: Proceed-
ings of 19th International Symposium on Artificial Life and Robotics, pp. 342–245
(2016)
4. Yamaba, H., Kurogi, T., Kubota, S., et al.: Evaluation of feature values of surface
electromyograms for user authentication on mobile devices. Artif. Life Robot. 22,
108–112 (2017)
5. Yamaba, H., Kurogi, T., Aburada, A., et al.: On applying support vector machines
to a user authentication method using surface electromyogram signals. Artif. Life
Robot. (2017). https://doi.org/10.1007/s10015-017-0404-z
6. Kurogi, T., Yamaba, H., Aburada, A., et al.: A study on a user identification method
using dynamic time warping to realize an authentication system by s-EMG. Adv.
Internet Data Web Technol. (2018). https://doi.org/10.1007/978-3-319-75928-9 82
7. Yamaba, H., Aburada, A., Katayama, T., et al.: Evaluation of user identification
methods for an authentication system using s-EMG. Adv. Netw.-Based Inf. Syst.
(2018). https://doi.org/10.1007/978-3-319-98530-5 64
8. Kita, Y., Okazaki, N., Nishimura, H., et al.: Implementation and evaluation of
shoulder-surfing attack resistant users (in Japanese). IEICE Part D J97–D(12),
1770–1784 (2014)
9. Kita, Y., Kamizato, K., Park, M., et al.: A study of rhythm authentication and
its accuracy using the self-organizing maps (in Japanese). Proc. DICOMO 2014,
1011–1018 (2014)
App-Collusion Detection Using
a Two-Stage Classifier
Md. Faiz Iqbal Faiz(B) , Md. Anwar Hussain, and Ningrinla Marchang
Abstract. Various single app analysis tools are developed for checking
Android malwares in the smartphones. All of them are unable to detect
threats caused by more than one app. Android app-collusion is one such
threat. App-collusion is a scenario where two or more apps collaborate
with each other to achieve a malicious goal. This paper presents an app-
roach for detecting collusive app-pairs using machine learning. The pro-
posed approach works in two stages. In the fist stage, we train a base
classifier using a set of benign and malicious applications. In the second
stage, we use the parameter vector from the first stage and a different
classifier to detect collusive app-pairs. We achieve detection rates of 90%
and 87.5% on two sets of colluding app-pairs.
1 Introduction
Android employs permission-based security mechanism. In this mechanism, data
of individual apps are protected from unauthorized access. Each app is installed
with permissions. These permissions are requested at the time of installation.
The permissions protect data of individual apps from other apps. We can say that
the permission-based security mechanism protects private data of users. During
app installation, if the requested permissions are not granted by the user then
the installation is cancelled (for devices running Android versions below 6.0).
But for versions above 6.0, the apps can ask for permissions during run-time.
The users can accept or deny the requested permission.
Individual apps are executed in separate sandboxes. Protecting sensitive
resources from unauthorized access by the third-party apps. The Android operat-
ing system cannot detect if an app accessing a permission-protected resource via
another app has itself requested the permission. An app can have access to sensi-
tive resources which are permission protected and can limit access to its data by
using permissions. The developer of the app can expose the permission-protected
resources advertently or inadvertently. The applications can collaborate with
each other to gain access to permission-protected resources. Android malware
tools are capable of analysing single apps. However, they cannot detect threats
that span multiple apps. The colluding apps have permissions which when com-
bined, can bypass any security rule. The individual apps in the colluding app
set may appear benign and incapable of executing a threat [1,4,6,8,11,12].
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 702–710, 2019.
https://doi.org/10.1007/978-3-030-15035-8_68
App-Collusion Detection Using a Two-Stage Classifier 703
2 Related Work
The idea of app collusion is first explained in [10], in which it is demonstrated
that Android’s permission-based mechanism is incapable of preventing appli-
cation collusion attacks. The authors showed using throughput measurements
of communication channels that even covert channels with low throughput can
leak private data out of smartphone. They posited app collusion is an open prob-
lem for the research community. The authors in [4] proposed XManDroid which
is an extension of Android permission framework to tackle privilege escalation
attacks. It analyses communication links between applications and ensures that
they comply to the desired system policy. The authors in [9] gave the first col-
lusion threat intelligence framework to detect app collusion. The system was
based on permission-based features and used Naive Bayes classifier with infor-
mative prior to calculate collusion threat. They also presented extensive analysis
of the role of permissions in collusion threat estimation. App-collusion detection
using informative logistic regression classifier is presented in [7], in which authors
discuss impact of various permissions.
3 Methodology
In this work, we present an app-collusion detection system for detecting app-
pairs which are colluding. We discuss app-collusion in Sect. 3.1, app-collusion
detection in Sect. 3.2, and finally the system model in Sect. 3.3.
3.1 App-Collusion
App-collusion is a scenario in which two or more apps secretly communicate with
each other and carry out a threat. The threat can be in the form of information
704 Md. F. I. Faiz et al.
theft, financial theft, service misuse, and elevation of privilege [15]. Information
exfiltration is one of the most prominent threat. Information leakage can be
carried out using the Internet feature or external storage of the device.
App-collusion is a multiple-app attacking model. The operations needed to
execute a threat are divided among multiple apps. However, when we look at
a single app, it appears benign. Each app can execute a portion of the given
operations using its permissions. When we look at it collectively, the colluding
app set includes dangerous combination of permissions capable of achieving a
malicious goal.
We give a formal definition of app collusion as in [7]. Assume A to be the set
of all android apps and P be the set of all dangerous permissions. Let a and b
be two applications with permission sets Pa and Pb such that a, b ∈ A and Pa ,
Pb ∈ P .
Now, assume that a performs a sensitive operation to access resource τ by
using permission β such that β ∈ (Pa −Pb ). If f (τ ) is a set of sequence of sensitive
operations on/using τ and if b performs a sequence of operations α ⊆ f (τ ) using
permission γ ∈ Pb . Then, this scenario is called collusion.
In this work, the number of apps in the colluding set is limited to two. But this
can be easily extended to more than two apps. We first train a machine learning
classifier using a corpus of benign and malicious applications. The features for
the classifier are critical permissions which are frequently requested by Android
malwares. We call this the base classifier for our detection system. More details
are provided in Sect. 3.3. We assume that threats posed by colluding apps are
no different than those posed by single apps.
After the base classifier is trained with a set of benign and malicious appli-
cations, we extract the learned parameters of each feature. The output of the
base classifier is the parameter vector. Each parameter in the parameter vector
corresponds to a feature of the data samples. The resultant parameter vector is
fed to a final classifier. This classifier takes the learned parameter vector and an
app-pair as input. The output of the classifier is a decision signalling whether
the given app-pair is colluding or non-colluding.
In the next section, we explain in detail our system model and the classifiers
used in it.
The app-collusion detection system consists of two main components. The first
component is the base classifier and the second component is the final classifier.
In this model, we use naive Bayes classifier as the base classifier. The base
classifier is trained with a set of benign and malicious applications. The features
we use are the critical permissions. These permissions are frequently requested by
Android malwares and some fall under dangerous category of permissions. We get
App-Collusion Detection Using a Two-Stage Classifier 705
the parameters that is calculated by training the base classifier. Each parameter
corresponds to a feature. So, the number of parameters is equal to the number of
features. These parameters are used by the final classifier for detecting collusive
app-pairs. The rationale behind using naive Bayes classifier as the base classifier
and a simple final classifier is that they are not compute-intensive compared
to other machine learning algorithms such as neural networks. We know that
the smartphones have limited amount of computational resources, hence using
low compute-intensive and simple classifier is the best choice. In the following
section, we explain the two main components.
Where cnb represents the target value output by the naive Bayes classifier.
Here, P (xj |cl ) are the feature probabilities conditioned on the classes. Detailed
explanation on naive Bayes classifier is found in [5]. After training the classifier
with the training set, we get the parameter vector θl = (θ1l , . . . , θkl ) for each class
cl where each parameter: θjl = P (xj |cl ). We use the parameter vector θ in the
final classifier.
4 Experimental Results
In this experiment, we train our base classifier with a training set of 5000 benign
and 3000 malicious applications. We test our base classifier on a set of 2000
benign and 207 malicious applications. We further test our base classifier on
three sets of malicious applications which consist of 1260, 247, and 154 malicious
applications obtain from [13]. We design two feature sets. The first set consists of
13 critical permissions which are requested by both Android malwares and col-
luding app-pairs. The second feature set consists of permissions used in first set
plus 13 extra permissions which are frequently requested by Android malwares.
The parameter vector is obtained as a result from training our base classifier.
The parameter vector is used by the final classifier to identify an app-pair as
malicious or non-malicious as explained in the previous section. For testing of
app-pairs, we use two sets of 120 colluding app-pairs obtain from [2].
Figures 1 and 2 shows the TPR plot of the base classifier on the four sets of
malwares. In this experiment, set 1, set 2, set 3, and set 4 consists of 207, 1260,
247, and 154 malicious applications respectively. The base classifier is trained
with feature sets of size 13 and 26. In other words, we design two permission
sets of size 13 and 26 as feature sets for training our base classifier. True positive
rates (TPR) for all the four sets have considerably improved when we go from
feature set size 13 to 26 (Table 3). It is expected because we have included a
greater number of critical permissions. We have used only the parameter vector
obtain from training our classifier with a feature set of size 13 because colluding
app-pairs requests only these permissions. This parameter vector is used in the
identification of colluding app-pairs.
App-Collusion Detection Using a Two-Stage Classifier 707
Permission name
GET TASKS
CHANGE WIFI STATE
RECORD AUDIO
READ PHONE STATE
READ HISTORY BOOKMARKS
READ CONTACTS
INTERNET
WRITE EXTERNAL STORAGE
READ EXTERNAL STORAGE
ACCESS WIFI STATE
GET ACCOUNTS
ACCESS FINE LOCATION
READ CALL LOG
Figures 3, 4, 5 and 6 show the TPR plots for our final classifier on the two
sets of colluding app-pairs. We have used two permission sets A and B as feature
sets. Figures 3 and 4 shows the effects of removing permission set A on the TPR
plot of the final classifier. We train the final classifier first with permission set
A only and then with permission set obtain by removing set A from the total
feature set. The TPR considerably increases for both the sets when we use set
A. Same is the case for permission set B as shown in Figs. 5 and 6.
Figures 7 and 8 show the TPR plots for our final classifier when we remove
two important critical permissions GET TASKS and INTERNET. As we can
see, the TPR decreases considerably when we remove these permissions. Tables 1
and 2 shows the list of permissions in the set A, set B, and the full feature or
permission set. Figure 9 shows the TPR plot when we use the full feature set.
We can see that the TPR decreases considerably. The reason behind this is that
some of the permissions or features are redundant or contribute negatively to
the decision. When we remove those features and train the final classifier with
permission set A or permission set B, the TPR increases considerably.
708 Md. F. I. Faiz et al.
Fig. 1. Base classifier TPR plot 1 Fig. 2. Base classifier TPR plot 2
Fig. 3. Final classifier TPR plot 1 Fig. 4. Final classifier TPR plot 2
Fig. 5. Final classifier TPR plot 3 Fig. 6. Final classifier TPR plot 4
Fig. 7. Final classifier TPR plot 5 Fig. 8. Final classifier TPR plot 6
App-Collusion Detection Using a Two-Stage Classifier 709
5 Conclusion
In this work, we propose a novel and simple technique for detecting collusive app-
pairs. We use a two-tier model. In the first step, we train a classifier for detecting
malicious applications using critical permissions as features. In the next step, we
use another classifier for detecting collusive app-pairs. We use the idea that a
collusive app-pair can carry out threats like the malicious applications. The final
classifier uses the parameter vector obtain from the first step for the detection.
The app-collusion detection system achieves a TPR of 90% and 87.5% on two
sets of colluding app-pairs. This can improve by including different kinds of
attributes as features. For instance, api calls can serve another set of features
which we can employ in our final classifier. Our approach is scalable and can be
used for large scale classification of app-pairs.
Acknowledgements. This work was funded by grant from the Visvesvaraya Fellow-
ship Scheme (Govt. of India).
References
1. Bartel, A., Bissyande, T., Klein, J., Traon, Y.L., Azert, S., Li, L.: IccTA: detecting
inter-component privacy leaks in android apps. In: Proceedings of 37th Interna-
tional Conference on Software Engineering, Florence, IEEE, pp. 280–291 (2015)
2. Blasco, J., Chen T.M.: Automated generation of colluding apps for experimental
research. J. Comput. Virol. Hack. Tech. 14, 127 (2018). https://doi.org/10.1007/
s11416-017-0296-4
3. Bosu, A., Liu, F., Yao, D., Wang, G.: Collusive data leak and more: large-scale
threat analysis of inter-app communications. In: Proceedings of Asia Conference
on Computer and Communications Security, pp. 71–85. ACM (2017)
4. Bugiel, S., Davi, L., Dmitrienko, A., Fischer, T., Sadeghi, A.R.: XManDroid: a
new android evolution to mitigate privilege escalation attacks. Technical report
TR -2011-04, Technische Universitat Darmstadt (2011)
5. Bishop, C.: Pattern Recognition and Machine Learning, pp. 1–758. Springer, New
York (2006)
710 Md. F. I. Faiz et al.
6. Chin, E., Porter, A.F., Greenwood, K., Wagner, D.: Analyzing inter-application
communication in Android. In: Proceedings of the 9th International Conference on
Mobile systems, Applications, and Services, pp. 239–252. ACM (2011)
7. Faiz, M.F.I., Hussain, A.M., Marchang, N.: Detection of collusive app-pairs using
machine learning. In: Proceedings of 3rd International Conference on Consumer
Electronics ASIA, pp. 206–212. IEEE (2018)
8. Felt, A.P., Wang, H.J., Moshchuk, A., Hanna, S., Chin, E.: Permission re-
delegation: attacks and defences. In: Proceedings of the 20th USENIX Security
Symposium. USENIX Association (2011)
9. Kalutarage, H.K., Nguyen, H.N., Shaikh, S.A.: Towards a threat assessment frame-
work for apps collusion. Telecommun. Syst. 66, 417–430 (2017)
10. Marforio, C., Ritzdorf, H., Francillon, A., Capkun, S.: Analysis of the communica-
tion between colluding applications on modern smartphones. In: Proceedings of the
28th Annual Computer Security Applications Conference, pp. 51–60. ACM (2012)
11. Octeau, D., Jha, S., Derring, M., McDaniel, P., Bartel, A., Li, L., Klein, J.,
Traon, Y.L.: Combining static analysis with probabilistic models to enable market-
scale android inter-component analysis. In: Proceedings of 43rd Annual ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pp.
469–484. ACM (2016)
12. Porter, A.F., Ha, E., Egelman, S., Haney, A., Chin, E., Wagner, D.: Android per-
missions: user attention, comprehension, and behaviour. In: Proceedings of the 8th
Symposium on Usable Privacy and Security, Article 3, pp. 1–14. ACM (2012)
13. Wang, W., Wang, X., Feng, D., Liu, J., Han, Z., Zhang, X.: Exploring permission-
induced risk in android applications for malicious application detection. IEEE
Trans. Inf. Forensics Secur. 9(11), 1869–1882 (2014). https://doi.org/10.1109/
TIFS.2014.2353996
14. Xie, Z., Zhu, S.: GroupTie: toward hidden collusion group discovery in app stores.
In: Proceedings of Conference on Security and Privacy in Wireless & Mobile Net-
works, pp. 153–164. ACM (2014)
15. Intel newsroom. https://newsroom.intel.com/news-releases/mcafee-labs-report-
reveals-new-mobile-app-collusion-threats/
The 5th International Workshop on
Engineering Energy Efficient
InternetWorked Smart seNsors
(E3WSN-2019)
Comparison of Machine Learning
Techniques for Prediction Problems
1 Introduction
Owning to the fact that considering the improvements in the artificial intelli-
gence in last two decades are affecting the multi-disciplinary fields electronically.
Within these developments many different and interesting proposals, frameworks
and models are available in the literature. Especially, back propagation neural
networks (BPNN) with prediction are highly acceptable and deployments have
been increased especially in the fields of healthcare systems, finance and banking,
agriculture, petroleum and gas and many more [1–6].
In their paper [1], authors suggested a gradient-based BPPN model for
improving optimization in stock price predictions. They used gradient as descent
in BPNN model to decide the parameter of learning rate, for training cycle in
order to find out the best value in the process of stock data training for obtain-
ing accuracy in prediction. According to their findings stock predictions and
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 713–723, 2019.
https://doi.org/10.1007/978-3-030-15035-8_69
714 Y. Kirsal Ever et al.
Please note that the first line of text that follows a heading is not indented,
whereas the first lines of all subsequent paragraphs are.
Radial Basis Function Neural Networks was proposed in 1988 [11] and basically
it uses radial basis functions as an activation function of hidden neurons. It
acts as universal approximation on sets and its’ strength areas are function
approximation, time series prediction and classification [12,21].
where n is the total number of samples and Yi and Ŷi are the predicted and
expected outputs of estimator.
R2 Score is variance of predictable sample from the independent sample. It
is defined as:
EVs
R2 = (3)
U Vs
718 Y. Kirsal Ever et al.
3.2 Results
As it is mentioned above, in this research, 8 datasets are considered to analyse
and compare 4 machine learning algorithms. Table 1 shows obtained results for
all datasets and machine learning techniques.
In Air Quality Dataset, Decision Tree Regressor and BP produced optimum
results when we consider all evaluation methods. SVR produced close results to
DTR and BP in R2 and EV Scores but MSE result is relatively higher than the
others. RBFNN produced inefficient results for all evaluation metrics. Figure 1
shows prediction graph of DTR for Air Quality Dataset.
In Wine Quality Dataset, DTR was unable to predict features correctly and
marked as not applicable both for Red and White group. Similarly, BP and SVR
Comparison of Machine Learning Techniques for Prediction Problems 719
produced close but insufficient results for each other, and optimum results are
obtained by RBFNN with a big variance. Figure 2 shows minimum prediction of
WQ dataset White group by BP.
In CCPP dataset, closer results are obtained for each machine learning tech-
nique in all shuffled datasets. BP produced optimum results for all in MSE, R2
Score and EV Score. It is followed by SVR and DTR in all shuffled datasets and
RBFNN produced insufficient prediction rates. Figure 3 shows the prediction
graph of backpropagation neural network for CCPP - 2.
results when EV and R2 Scores are considered. Figure 4 shows the prediction
graph of SVR that obtained minimum MSE in RE dataset.
4 Discussions
5 Conclusions
Machine Learning Techniques are proposed to solve all kind of problems but each
technique may produce higher rates in different problem domains and datasets.
Several researches were conducted and concluded different results for this.
In this paper, preliminary experiments were performed using nine datasets
for prediction in different kind of problems and four frequently used machine
learning techniques are considered for the comparison. Obtained results were
analyzed using three evaluation methods as MSE value, EV and R2 scores.
Obtained results showed that, incrementation or degradation of instances in
datasets do not affect the performance of the techniques directly but the charac-
teristics of datasets can affect the rates directly. When the analyses of evaluation
methods performed for the optimum results obtained, it can be concluded that
neural network algorithms, which are back-propagation and radial basis func-
tion neural network in this research, produced higher rates and steady results
than decision tree regressor. Similarly, support vector regression produced steady
results but could not produce optimum rates for datasets except mean square
error for real estate evaluation dataset.
Future work will include considering new and more datasets with higher
number of instances and attributes and performing comparison by adding other
techniques beside these, such as linear regression, long-short term memory etc.
Also the training and testing ratios will be investigating by degrading training
ratio to 60%.
722 Y. Kirsal Ever et al.
References
1. Utomo, D., Pujiono, Soeleman, M.A.: Stock price prediction using back propa-
gation neural network based on gradient descent with momentum and adaptive
learning rate. J. Internet Bank. Commer. 22, 1–16 (2017)
2. Nawi, N.M., Ghazali, R., Salleh, M.N.M.: The development of improved back-
propagation neural networks algorithm for predicting patients with heart disease.
In: Proceedings of Information Computing and Applications - First International
Conference, ICICA 2010, pp. 317–324 (2010)
3. Gu, J., Yin, G., Huang, P., Guo, J., Chen, L.: An improved back propagation
neural network prediction model for subsurface drip irrigation system. Comput.
Electr. Eng. 60, 58–65 (2017)
4. Lee, T.-L.: Back-propagation neural network for the prediction of the short-term
storm surge in Taichung harbor, Taiwan. Eng. Appl. Artif. Intell. 21(1), 63–72
(2008)
5. Ogunde, A.O., Ogunleye, G.O., Oreoluwa, O.: A decision tree algorithm based
system for predicting crime in the university. Mach. Learn. Res. 2(1), 26–34 (2017)
6. Mesaric, J., Šebalj, D.: Decision trees for predicting the academic success of stu-
dents. Croat. Oper. Res. Rev. 7, 367–388 (2016)
7. Lee, J.S., Lee, E.S.: Exploring the usefulness of a decision tree in predicting people’s
locations. Procedia - Soc. Behav. Sci. 140, 447–451 (2014)
8. Gupta, B., Arora, A., Rawat, A., Jain, A., Dhami, N.: Analysis of various decision
tree algorithms for classification in data mining. Int. J. Comput. Appl. 163(8),
15–19 (2017)
9. Tuncal, K., Sekeroglu, B., Ozkan, C.: Lung cancer incidence prediction using
machine learning algorithms. Accepted to be published in International Journal
of Machine Learning and Computing (2019). Student Performance Prediction and
Classification Using Machine Learning Algorithms
10. Sekeroglu, B., Dimililer, K., Tuncal, K.: Student performance prediction and clas-
sification using machine learning algorithms. In: Accepted to be published in Pro-
ceedings of 8th International Conference on Educational and Information Technol-
ogy, ICEIT 2019 (2019)
11. Broomhead, D.S., Lowe, D.: Radial basis functions, mutli-variable functional inter-
polation and adaptive networks. Complex Syst. 2, 321–355 (1988)
12. Rout M., Majhi, B., Mohapatra, U.M., Mahapatra, R.: Stock indices prediction
using radial basis function neural network. In: Proceedings of International Con-
ference on Swarm, Evolutionary, and Memetic Computing, pp. 285–293 (2012)
13. Smola, A.J., Scholkopf, B.: A tutorial on support vector regression. Stat. Comput.
14(3), 199–222 (2004)
14. Henrique, B.M., Sobreiro, V.A., Kimura, H.: Stock price prediction using support
vector regression on daily and up to the minute prices. J. Financ. Data Sci. 4(3),
183–201 (2018)
15. Azeez, O.S., Pradhan, B., Shafri, H.Z.M.: Vehicular CO emission prediction using
support vector regression model and GIS. Sustainability 10(10), 3434 (2018)
16. Dougherty, G.: Pattern Recognition and Classification, 1st edn. Springer, New York
(2013)
17. De Vito, S., Massera, E., Piga, M., Martinotto, L., Di Francia, G.: On field calibra-
tion of an electronic nose for benzene estimation in an urban pollution monitoring
scenario. Sens. Actuators B: Chem. 129(2), 750–757 (2008)
Comparison of Machine Learning Techniques for Prediction Problems 723
18. Zhong, P., Fukushima, M.: Regularized nonsmooth Newton method for multi-class
support vector machines. Optim. Methods Softw. 22(1), 225–236 (2007)
19. Tüfekci, P.: Prediction of full load electrical power output of a base load operated
combined cycle power plant using machine learning methods. Int. J. Electr. Power
Energy Syst. 60, 126–140 (2014)
20. Yeh, I.C., Hsu, T.K.: Building real estate valuation models with comparative app-
roach through case-based reasoning. Appl. Soft Comput. 65, 260–271 (2018)
21. Ogidan E.T., Dimililer K., Ever Y.K.: Machine learning for expert systems in
data analysis. In: 2nd International Symposium on Multidisciplinary Studies and
Innovative Technologies, ISMSIT 2018 (2018)
Tailoring Micro-solar Systems
to Heterogeneous Wireless Sensor
Networks
1 Introduction
Wireless sensor networks (WSNs) are a fundamental part of what we call Inter-
net of Things. Through WSNs we are able to generate data we have never been
able to create before, both in terms of quality and quantity. Unfortunately, to
achieve the most from this technology we need these networks to be deployed
wherever the physical phenomenon to monitor happens, and we often need them
to run for long periods. These factors lead researchers to consider energy one
of the key constraints in WSNs [11]. Clustering algorithms have been intro-
duced in WSNs to address this challenge. This approach organises the devices
into sets (clusters). Each of them has a cluster head that gathers data from its
nodes and communicates with other cluster heads in order to report data to
a centralised data sink [10]. While clustering optimises nodes energy consump-
tion, engineers introduced energy harvesting to provide wireless sensor nodes
more energy. Micro-solar power systems are one of the most used harvesting
techniques [8], due to the need to equip nodes with very compact components.
Thanks to micro-solar power systems nodes can recharge their batteries using
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 724–733, 2019.
https://doi.org/10.1007/978-3-030-15035-8_70
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 725
solar power, thus increasing network lifetime. Many efforts have been put in this
field in order to get to the point where the network is able to perform its task
with harvesting as its only energy supply: this result is called energy neutrality
[3,4]. As we stress in Sect. 2, researchers achieved energy neutrality through ad-
hoc designs. This poses constraints and limitation to WSN developers, in terms
of hardware, software, number of nodes and network structure. To mitigate these
limitations we propose a new approach to energy harvesting WSN design. We
decide to base our work on widely-known and accepted clustering protocols for
WSNs. This way, we achieve the best results in terms of energy consumption
while maintaining a general purpose network architecture. Then, we study the
energy supply the network needs to reach energy neutrality and we derive a
suitable micro-solar power system. In Sect. 3 we define four different network
configurations and we simulate REECHD protocol on them to measure their
energy needs. Then, in Sect. 4 we use such information to design a solar harvest-
ing system with the minimal dimensions of batteries and solar panels needed
to achieve energy neutrality. In Sect. 5, we show how much energy the network
needs to harvest to reach energy neutrality in the four networks aforementioned,
as well as the necessary hardware. Finally, in Sect. 6 we show the implications
of the result we achieved, and we propose some future works.
2 Related Works
In this section we analyse the current state of the art in terms of clustering
algorithms, energy harvesting techniques, and the combination of these 2 features
in WSNs.
between neighbour sensor nodes. Moreover, if a node gets the CH role, it can’t
be re-elected as CH in the next cluster election. The communication between
CHs and the BS is single-hop.
HEED [17] is another well-known clustering protocol which aim to produce
clusters with the same size. HEED uses node residual energy to perform CH
election in the election phase. Then, neighbour nodes become member nodes
by joining the least costly (usually the closest) CH. In the steady phase each
member node sends a message to its CH. Then, CHs forward the messages to
the BS. It is worth to mention that HEED algorithm prevents two nodes within
the same transmission radius to simultaneously become CHs. HEED introduces
the notion of “round” as a grouping of subsequent TDMAs. A TDMA is the
time between a node starting to collect data and the data reaching the BS.
A problem which affects HEED is the cost of the CHs election phase, which
can decrease the lifetime of the nodes by depleting their energy too fast. ER-
HEED [15] aims to improve the performance of HEED by rotating the CHs role
inside clusters. Clusters are initially formed through a classical HEED election
and formation phase, while in the next rounds each CH choose the next CH as the
cluster member node with the highest energy left. HEED cluster head election
has to be repeated only if a sensor node inside a cluster depletes its battery
completely. ER-HEED rotation algorithm makes it more effective in terms of
network lifetime of HEED.
Rotated Unequal HEED (RUHEED) [1] produces unequal-sized clusters by
decreasing the competition radius of nodes near to the BS, thus reducing the
hot-spot problem. Moreover, it adds a rotation phase after cluster election and
formation. Rotation occurs when the current CH designates the next CH among
its cluster members basing on their residual energy. The node with the highest
residual energy becomes the next CH. Rotation avoids re-election of the CHs,
reducing the communication overhead and thus postponing the death of the net-
work. Cluster election and cluster formation take place when any of the sensors
dies.
DEEC (distributed energy-efficient clustering algorithm for heterogeneous
WSNs) [12] is a clustering protocol which takes into account the energy het-
erogeneity of the network. In fact, the probability of each node to become CH
is based on the ratio between the residual node energy and the whole network
average energy. Thus, nodes with both high residual and initial energies have
more chances to become CHs.
REECHD [10] is another clustering algorithm for heterogeneous WSNs. A
peculiarity of REECHD is that it considers the transmission rate of the nodes in
the CH leader election. In particular, the election process considers the average
transmission rate of the node to estimate the extent of work it induces on its
neighbours. In fact, nodes with higher transmission rates induce less work on
neighbours and thus they should have a higher probability of becoming CH. On
the other hand, nodes with low transmission rates should have less probability of
becoming CH. That is because they generate little intra-traffic communication
inside their cluster when they are not CH.
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 727
3 Scenario
We want to measure the energy we need to harvest to reach energy neutrality
on four different network configuration. To do this, we use the REECHD [10]
algorithm to cluster each network then for every node we get the maximum
and average energy consumption over all the Rounds until first node dies. The
energy depleted is the energy needed by each node for communication, network
synchronisation, and data collection, aggregation and forwarding to the BS.
We choose REECHD because it is one of the most optimized algorithms, and
because it is based on the well known MIT radio model [7]. The MIT model cal-
culates the transmission and reception energy consumed in sending and receiving
a data packet of k bits over a distance d as follows:
ET x (k, d) = k(Eelec + Ea dn ) (1)
ERx (k) = k(Eelec ) (2)
where Eelec = 50 nJ/bit is the energy consumption of a sensor transceiver cir-
cuit and Ea is the sensor amplification energy which depends on the distance
between a sender node and a receiver. When d < d0 = 75 m, Ea becomes Ef s =
10 pJ/bit/m2 while when d ≥ d0 = 75 m Ea becomes Emf = 0.0013 pJ/bit/m4 .
First, we define four different networks where nodes are deployed on a fixed
size square area. Nodes are not mobile i.e. they are stationary, while the BS is
located at the centre of the area.
• hW SN1 Area 100 m × 100 m, 100 nodes;
• hW SN2 Area 200 m × 200 m, 100 nodes;
• hW SN3 Area 500 m × 500 m, 100 nodes;
• hW SN4 Area 1000 m × 1000 m, 100 nodes.
We vary the Area size and keep the number of nodes fixed, to lower the density
of the network and increase the average distance between the nodes. Moreover,
we consider two different scenarios for each area: one in which the data sensed
from the sensors is not aggregated when sent to other nodes toward the BS, and
another where such data is aggregated. An aggregation value equal to 0 means
that all messages are sent to the BS, whereas an aggregation value of 1 results
in only one message sent to the BS (due to data compression).
In hW SN1 nodes are uniformly distributed inside the network, while in
hW SN2 , hW SN3 and hW SN4 the same nodes are deployed after different lin-
ear changes of coordinates. Due to the large size of the networks, instead of
implementing the REECHD protocol on top of hardware sensors we use a Java
based simulation tool specifically designed to run clustering protocols in hWSNs
networks. The parameters used to simulate the four scenarios described at the
top of this section are listed in Table 1.
After the simulation, we have all the data we need to define the minimum
hardware requirements as shown in the next section. Table 2 shows the energy
depleted by the nodes of network hW SN1 with no aggregation and nodes com-
petition radius set to 25 m. Emax is the maximum energy while Eaverage is the
average energy depleted by that node for each Round.
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 729
Simulation parameters
Parameters Values
Network grid 100 × 100, 200 × 200
500 × 500, 1000 × 1000
BS (50, 50), (100, 100)
(250, 250), (500, 500)
Eelec 50 nJ/bit
Ef s 10 pJ/bit/m2
Emp 0.0013 pJ/bit/m4
Homogeneous Tx rate 2000
Homogeneous energy 1J
Heterogeneous Tx rate (500...2000)
Heterogeneous energy (1...4 J)
R0 25 m, 50 m, 125 m, 250 m
Number of nodes 100
Aggregation rate 0,1
into each battery is computed as the product of En daily and 7 divided by the
output regulator efficiency (in our case, 50%), divided by the maximum dis-
charge percentage for the batteries (fixed to 80% of initial capacity) as shown in
Fig. 1. Finally, we can compute the minimum amount of energy the solar panel
have to output in order to achieve the neutral energy consumption goal. This is
the ratio between the energy to store into each battery and the input regulator
efficiency (fixed to 70%) divided by the batteries conversion efficiency which is
equal to 99%.
5 Results
In this section we show the results of our study. To get our results, we start from
existing climatic historical data available on the internet1 . We need this data to
estimate the average number of sunny days over the months of one year.
It is worth mentioning that we measure the minimum hardware require-
ments starting from the maximum energy needs we get from the simulations
(see Sect. 3). In fact, not all the nodes have the same energy needs. Nodes near
the BS, for instance, usually deplete their energy faster because they have to
forward data from the lower levels of the network. So, considering the highest
energy needs we get from each simulation, we can ensure that every node receives
a sufficient energy input from its harvester (i.e. the solar panel).
By looking at Table 3, we observe that both 100 × 100 and 200 × 200 scenarios
have similar energy needs in order to reach neutral energy consumption. In partic-
ular, they need a 10 cm × 10 cm panel with a peak power of 1 Wp both when the
aggregation is equal to 1 (A) and when it is equal to 0 (NA). Differently, if we con-
sider the 500 × 500 scenario we can appreciate a big difference between the A and
NA cases. In fact, the former is way less energy-hungry of the latter, which needs
a three times bigger solar panel and eight times more batteries to achieve neutral
energy consumption. Finally, in the 1000 × 1000 case, we can see that we need a
1
http://re.jrc.ec.europa.eu/pvg tools/en/tools.html.
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 731
solar panel of 1,16 m2 area and a 24 batteries pack for each sensor. This makes
clear that the long distances between the nodes affects the energy need in dispro-
portionate way, raising the minimum energy harvesting requirements to 31,78 Wh.
It is clear that the amount of harvested energy is much more than the actual
need of the nodes. One of the reasons is that there are times of the day with no
or few energy output from the harvester (i.e. at night). The graphs in Figs. 2, 3,
4 and 5 show the gap between the energy output and the energy not captured.
Table 3. Solar panels area needed for each node, according to the network config-
uration. A denotes an aggregation rate of 1, NA denotes an aggregation rate of 0.
Fig. 2. 100 × 100 with an aggregation Fig. 3. 200 × 200 with an aggregation
rate of 0. rate of 0.
Fig. 4. 500 × 500 with an aggregation Fig. 5. 1000 × 1000 with an aggregation
rate of 0. rate of 0.
732 S. Calabrò et al.
6 Conclusions
In this work, we studied the energetic needs of four different hWSNs clustered
with the same algorithm (REECHD) in order to compute the amount of energy
to harvest so that they could reach an energy neutrality state. To do this, we
found the actual energy consumed from each node during the network life cycle
thus we used such information to determine the minimum amount of energy
to harvest for reaching our goal. As a future work, we propose to extend our
approach to different energy harvesting sources (piezoelectric, wind, ... ), and
to measure performances on testbeds with real hardware and advanced software
frameworks as [5,13] or [6]. Moreover, we want to collect enough data from other
scenarios as a base to develop a solid analytical model able to calculate hWSNs
energy harvesting needs for any given network.
References
1. Aierken, N., Gagliardi, R., Mostarda, L., Ullah, Z.: Ruheed-rotated unequal clus-
tering algorithm for wireless sensor networks. In: 29th IEEE International Con-
ference on Advanced Information Networking and Applications Workshops, AINA
2015 Workshops, Gwangju, South Korea, 24–27 March 2015, pp. 170–174 (2015)
2. Anisi, M.H., Abdul-Salaam, G., Idris, M.Y.I., Wahab, A.W.A., Ahmedy, I.: Energy
harvesting and battery power based routing in wireless sensor networks. Wireless
Netw. 23(1), 249–266 (2017)
3. Besbes, H., Smart, G., Buranapanichkit, D., Kloukinas, C., Andreopoulos, Y.: Ana-
lytic conditions for energy neutrality in uniformly-formed wireless sensor networks.
IEEE Trans. Wireless Commun. 12(10), 4916–4931 (2013)
4. Chen, W., Andreopoulos, Y., Wassell, I.J., Rodrigues, M.R.D.: Towards energy
neutrality in energy harvesting wireless sensor networks: a case for distributed
compressive sensing? In: 2013 IEEE Global Communications Conference (GLOBE-
COM), pp. 474–479, December 2013
5. Dulay, N., Micheletti, M., Mostarda, L., Piermarteri, A.: PICO-MP: de-centralised
macro-programming for wireless sensor and actuator networks. In: 2018 IEEE 32nd
International Conference on Advanced Information Networking and Applications
(AINA), pp. 289–296, May 2018
6. Gummadi, R., Gnawali, O., Govindan, R.: Macro-programming wireless sensor
networks using Kairos. In: Proceedings of First IEEE International Conference on
Distributed Computing in Sensor Systems, DCOSS 2005, Marina del Rey, CA,
USA, 30 June–1 July 2005, pp. 126–140 (2005)
7. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-efficient commu-
nication protocol for wireless microsensor networks. In: Proceedings of the 33rd
Hawaii International Conference on System Sciences, vol. 8, HICSS 2000, p. 8020,
Washington, DC, USA. IEEE Computer Society (2000)
8. Jeong, J., Jiang, X., Culler, D.: Design and analysis of micro-solar power systems
for wireless sensor networks. In: 2008 5th International Conference on Networked
Sensing Systems, pp. 181–188, June 2008
9. Li, Y., Shi, R.: An intelligent solar energy-harvesting system for wireless sensor
networks. EURASIP J. Wirel. Commun. Netw. 2015(1), 179 (2015)
Tailoring Micro-solar Systems to Heterogeneous Wireless Sensor Networks 733
10. Micheletti, M., Mostarda, L., Piermarteri, A.: Rotating energy efficient clustering
for heterogeneous devices (REECHD). In: Barolli, L., Takizawa, M., Enokido, T.,
Ogiela, M.R., Ogiela, L., Javaid, N. (eds.) 32nd IEEE International Conference on
Advanced Information Networking and Applications, AINA 2018, Krakow, Poland,
16–18 May 2018, pp. 213–220. IEEE Computer Society (2018)
11. Miorandi, D., Sicari, S., Pellegrini, F.D., Chlamtac, I.: Internet of things: vision,
applications and research challenges. Ad Hoc Netw. 10(7), 1497–1516 (2012)
12. Qing, L., Zhu, Q., Wang, M.: Design of a distributed energy-efficient clustering
algorithm for heterogeneous wireless sensor networks. Comput. Commun. 29(12),
2230–2237 (2006)
13. Stanford-Clark, A., Truong, H.L.: MQTT for sensor networks (MQTTS) specifica-
tions, October 2007
14. Taneja, J., Jeong, J., Culler, D.: Design, modeling, and capacity planning for micro-
solar power sensor networks. In: 2008 International Conference on Information
Processing in Sensor Networks (IPSN 2008), pp. 407–418, April 2008
15. Ullah, Z., Mostarda, L., Gagliardi, R., Cacciagrano, D., Corradini, F.: A compari-
son of heed based clustering algorithms – introducing er-heed. In: 2016 IEEE 30th
International Conference on Advanced Information Networking and Applications
(AINA), pp. 339–345, March 2016
16. Yang, L., Lu, Y.-Z., Zhong, Y., Wu, X.G., Yang, S.X.: A multi-hop energy neu-
tral clustering algorithm for maximizing network information gathering in energy
harvesting wireless sensor networks. In: Sensors (2015)
17. Younis, O., Fahmy, S.: Heed: a hybrid, energy-efficient, distributed clustering app-
roach for ad hoc sensor networks. IEEE Trans. Mob. Comput. 3(4), 366–379 (2004)
18. Yu, H., Yue, Q.: Indoor light energy harvesting system for energy-aware wireless
sensor node. Energy Procedia 16, 1027–1032 (2012). 2012 International Conference
on Future Energy, Environment, and Materials
19. Zhang, P., Xiao, G., Tan, H.: Clustering algorithms for maximizing the lifetime
of wireless sensor networks with energy-harvesting sensors. Comput. Netw. 57,
2689–2704 (2013)
Distributing Energy Consumption
in Multi-interface Series-Parallel
Networks
1 Introduction
As technology advances and hardware costs are reduced, very powerful devices
are available for a wide range of applications. Moreover, heterogeneous devices
The work has been supported in part by the European project “Geospatial based
Environment for Optimisation Systems Addressing Fire Emergencies” (GEO-SAFE),
contract no. H2020-691161, and by the Italian National Group for Scientific Compu-
tation (GNCS-INdAM).
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 734–744, 2019.
https://doi.org/10.1007/978-3-030-15035-8_71
Multi-interfaces on Serial-Parallel Networks 735
In this paper, we are interested in the so-called Coverage problem, see [19]. It
consists in finding the cheapest way to establish all the connections defined by an
input graph G, no matter the interface used to accomplish each connection, no
bandwidths requirements are provided. The problem only asks to ensure that for
each edge of G, there is a common active interface at its endpoints. The objective
is to minimize the overall cost of activation in the network. The problem has been
extensively studied even with respect to another interesting objective function
that concerns the minimization of the maximum cost on a single node, see [12].
Consistently with [4], here we consider one further constraint to the original
setting, in order to maintain under control the energy spent by single devices.
A parameter p is introduced with the meaning that no device can activate more
than p interfaces. That is, instead of finding the solution that minimizes the over-
all cost due to the activation of the interfaces along the whole network, or the
one that minimizes the maximum cost spent by a single node, the aim becomes
to minimize the overall cost subject to constraint p. This is very important for
practical purposes as it is a step towards the study of multi-interface networks in
distributed setting. Similarly to [10], this is a new approach to maintaining bal-
anced the energy consumption of the single nodes. This approach clearly helps
in prolonging the lifetime of the network. A possible context where our theoret-
ical results can be exploited may concern military applications, and specifically
the tactical networks provided with new software technologies [3,23]. In fact,
this type of networks suffers for the employing of resource-constrained mobile
devices. Clearly, one of the main resources that has to be preserved is the battery
power of each device, this implies that the suitable usage of the interfaces might
be crucial for prolonging the network lifetime.
Multi-interface networks have been extensively studied during the last years in a
variety of contexts, usually focusing on the benefits of exploiting multiple inter-
faces of each device. Many basic problems of standard network optimization are
reconsidered in such a setting [7,10], in particular, focusing on issues related to
routing [13] and network connectivity [9,15]. Combinatorial problems on multi-
interface wireless networks have been studied starting from [8]. That paper, as well
as [4,12,19], investigate the Coverage problem and its variants described above.
Connectivity issues have been addressed in [5,22]. The problem consists in
finding the cheapest way to ensure the connectivity of the entire network. In
other words, it aims to find at each node a subset of the available interfaces that
must be activated in order to guarantee a path between every pair of nodes in
G while minimizing the overall cost of the interfaces activated among the whole
network. As for the case of Coverage, another studied objective function is that
of minimizing the maximum cost paid on a single node [12]. The Connectivity
problem corresponds to a generalization of the well-know Minimum Spanning
Tree problem. In fact, costs are not on the edges but on the interfaces hold by
the nodes. It may happen that the same interface is used by a node to establish
Multi-interfaces on Serial-Parallel Networks 737
In this paper, we are interested in the Coverage problem with the further con-
straint that each node of the network can activate at most 2 interfaces. According
to [4], from now on the problem is denoted by CMI(2) and we consider the so-
called unbounded case where there is no bound on the number k of interfaces
available over all the network. Since in general it has been already proved in [4]
that the problem is NP-hard, it is worth investigating on specific graph topolo-
gies. Basic ones like paths, rings, trees, complete graphs and complete bipartite
graphs have been already addressed and optimally solved in [4]. Here we keep on
investigating on a wide graph class, namely the Series-Parallel graphs [14]. The
formal definition of such intriguing topology will be given in the next section.
Informally, Series-Parallel graphs are graphs with two distinguished nodes called
terminals, formed recursively by two simple composition operations. Such graphs
have been largely investigated in the last decades, mainly for their relationship
with series and parallel electric circuits. We provide a polynomial time dynamic
programming algorithm that optimally solves CMI(2) on Series-Parallel graphs.
1.3 Outline
The next section provides definitions and notation in order to formally describe
the CMI(p) problem.
Section 3 is the core of the paper, containing the formal definition of Series-
Parallel graphs and the resolution algorithm that works in polynomial time.
Finally, Sect. 4 contains conclusive remarks and a discussion of interesting
open problems.
2 Preliminaries
For a graph G, we denote by V its node set, by E its edge set. Unless otherwise
stated, the graph G = (V, E) representing the network is assumed to be undi-
rected, connected, and without multiple edges and loops. For each node v ∈ V ,
let deg(v) be the degree of v, and Δ be the maximum degree among all the
nodes. When considering the network graph G, we simply denote the number of
its nodes and edges by n and m, respectively. A global assignment of the inter-
faces to the nodes in V is given in terms of an appropriate interface assignment
function W , as follows.
Definition 1. A function W : V → 2{1,2,...,k} is said to cover graph G if W (u) ∩
W (v) = ∅, for each {u, v} ∈ E.
The cost of activating an interface for a node is assumed to be identical for all
nodes and given by a cost function c : {1, . . . , k} → R+ , i.e., the cost of interface
i is denoted as ci . The considered CMI(p) optimization problem is formulated
as follows.
Multi-interfaces on Serial-Parallel Networks 739
Note that we can consider two variants of the above problem: The cost func-
tion c can span over R+ , or c(i) = 1, with i = 1, . . ., k (unit cost case). In both
cases we assume k ≥ 2, since the case k = 1 admits an obvious unique solution
(all nodes must activate their unique interface).
In [4] (and similarly to the approach adopted in [1]), it has been shown that
CMI(p) is a special case of the more general CMI(∞) problem (see [10,19]),
where each node cannot activate more than p interfaces. Interestingly, on the
one hand the basic variant with p = 2 results more difficult in general than
CMI(∞). On the other hand, there are special graph classes that turn out to be
much more affordable. For trees and complete graphs, for instance, CMI(∞) has
been proved to be APX -hard and not approximable within O(log k), respectively,
while CMI(2) is polynomially solvable.
In particular, the subgraph G(v) (induced by the edges of the subtree rooted in
the interior node v) can be constructed by a series or a parallel composition of
the two SP-graphs G(v1 ) and G(v2 ) where v1 and v2 are the children of v in the
binary tree. The authors of [24] described an O(n) algorithm that gets a graph
G as input and returns a binary tree representation of the G if and only if the
G is an SP-graph. Moreover, it is well-known that the number m of edges of an
SP-graph satisfies the equations n − 1 ≤ m ≤ 2n − 3, and a simple proof can be
found in [20].
An example of an SP-graph, and a possible binary decomposition tree are
depicted in Figs. 1 and 2, respectively.
This section describes a polynomial time optimal algorithm for CMI(2) using
the dynamic programming technique. The algorithm exploits the binary tree
representation of an SP-graph provided above. Then, the algorithm solves the
subproblems on the leaves and then iteratively extends these solutions until the
root where the whole problem is solved.
In fact, to be a proper series composition, the node h is the sink in G(v1 ) and
the source in G(v2 ). So, the cost of the active interfaces in C associated to h is
counted twice. This is the reason why we subtract c(C) in the objective function.
Analogously, we can easily compute the optimal value of an interior node v
of T with label P (parallel composition) and children v1 and v2 . Let i and j
be the source and the sink, resp., of both G(v1 ) and G(v2 ), then the value of
f (v, A, B) is equal to:
As nodes i and j are the source and the sink, resp., of both G(v1 ) and G(v2 ),
then the costs of the active interfaces A and B are counted twice. This is the
reason why we subtract c(A) and c(B) in the objective function.
Using this algorithm, the optimal value of CMI(2) is the minimum value
f (root, A, B), for every possible subsets A and B of the available interfaces in
the source i and the sink j of the root of T , with |A| ≤ 2 and |B| ≤ 2. At
each node labelled with S, O((k + k2 )3 ) values f (v, A, B) are computed. In
fact, we need to test on three subsets A, B, and C of interfaces containing at
most 2 interfaces each,
and all their combinations. Instead, at each node labelled
with a P, O((k + k2 )2 ) values f (v, A, B) are computed as we need to test two
subsets A and B of interfaces containing at most 2 interfaces each, and all their
combinations. Since the number of the edges in an SP-graph is at most 2n − 3,
then, the numberofthe nodes in T is O(n). Summarizing, the dynamic algorithm
runs in O(n(k + k2 )3 ) time. We can then state the following:
Theorem 1. CMI(2) is solvable on Series-Parallel graphs in O(k 6 n) time.
742 A. Aloisio et al.
4 Concluding Remarks
In this paper, we have continued our investigation on a variant of the well-
studied Coverage problem arisen in the context of multi-interface networks. The
task asks to find the cheapest way to establish all the connections defined by an
input graph by activating suitable subsets of interfaces available at the nodes
of the network. The modification to the original model concerns the further
constraint for which each node can activate at most p interfaces. In particu-
lar we focus on the case with p = 2 and considering Series-Parallel graphs as
underlying topology. The problem is known to be NP-hard in general whereas
we could devise a polynomial time algorithm that optimally solve the problem
in the considered context. The new proposed algorithm is based on dynamic
programming.
Considering that Series-Parallel graphs admit a treewidth of at most 2, it would
be interesting to keep on investigating whether the problem can be solved in graphs
with bounded treewidth. Also planar graphs would be worth to investigate as
Series-Parallel graphs contain the class of outerplanar graphs. Moreover, as the
NP-hardness is given for general graphs of maximum degree 4, and that for max-
imum degree 2 the problem is optimally solvable in polynomial time, it remains
to investigate the case of graphs with maximum degree 3. Finally, new interesting
research lines in the context of Multi-Interface networks could be that of consid-
ering small-world [16] or greedy coloring [18] approaches, or multi-objective func-
tions. For instance, similar to [2], a requirement might be to minimize the number
of types of used interfaces subject to a bounded global cost.
References
1. Aloisio, A., Arbib, C., Marinelli, F.: Cutting stock with no three parts per pattern:
work-in-process and pattern minimization. Discrete Optim. 8, 315–332 (2011)
2. Aloisio, A., Arbib, C., Marinelli, F.: On LP relaxations for the pattern minimization
problem. Networks 57, 247–253 (2011)
3. Aloisio, A., Autili, M., D’Angelo, A., Viidanoja, A., Leguay, J., Ginzler, T., Lampe,
T., Spagnolo, L., Wolthusen, S., Flizikowski, A., Sliwa, J.: TACTICS: tactical ser-
vice oriented architecture. In: 3rd International Conference in Software Engineering
for Defence Applications, pp. 1–9 (2014)
4. Aloisio, A., Navarra, A.: Balancing energy consumption for the establishment of
multi-interface networks. In: Proceedings of the 41st International Conference on
Current Trends in Theory and Practice of Computer Science, (SOFSEM). LNCS,
vol. 8939, pp. 102–114. Springer (2015)
5. Athanassopoulos, S., Caragiannis, I., Kaklamanis, C., Papaioannou, E.: Energy-
efficient communication in multi-interface wireless networks. Theory Comput. Syst.
52, 285–296 (2013)
6. Audrito, G., Bertossi, A., Navarra, A., Pinotti, C.: Maximizing the overall end-user
satisfaction of data broadcast in wireless mesh networks. J. Discrete Algorithms
45, 14–25 (2017)
7. Bahl, P., Adya, A., Padhye, J., Walman, A.: Reconsidering wireless systems with
multiple radios. SIGCOMM Comput. Commun. Rev. 34(5), 39–46 (2004)
Multi-interfaces on Serial-Parallel Networks 743
8. Caporuscio, M., Charlet, D., Issarny, V., Navarra, A.: Energetic performance of
service-oriented multi-radio networks: issues and perspectives. In: Proceedings of
the 6th International Workshop on Software and Performance (WOSP), pp. 42–45.
ACM (2007)
9. Cavalcanti, D., Gossain, H., Agrawal, D.: Connectivity in multi-radio, multi-
channel heterogeneous ad hoc networks. In: Proceedings of the 16th International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), pp.
1322–1326. IEEE (2005)
10. D’Angelo, G., Di Stefano, G., Navarra, A.: Multi-interface wireless networks: com-
plexity and algorithms. In: Ramakrishnan, S., El Emary, I.M.M. (eds.) Wireless
Sensor Networks: From Theory to Applications, pp. 119–155. CRC Press/Taylor
& Francis Group, Boca Raton (2013)
11. D’Angelo, G., Di Stefano, G., Navarra, A.: Flow problems in multi-interface net-
works. IEEE Trans. Comput. 63, 361–374 (2014)
12. D’Angelo, G., Stefano, G.D., Navarra, A.: Minimize the maximum duty in multi-
interface networks. Algorithmica 63(1–2), 274–295 (2012)
13. Draves, R., Padhye, J., Zill, B.: Routing in multi-radio, multi-hop wireless mesh
networks. In: Proceedings of the 10th International Conference on Mobile Com-
puting and Networking (MobiCom), pp. 114–128. ACM (2004)
14. Duffin, R.: Topology of series-parallel networks. J. Math. Anal. Appl. 10(2), 303–
318 (1965)
15. Faragó, A., Basagni, S.: The effect of multi-radio nodes on network connectivity—
a graph theoretic analysis. In: Proceedings of the 19th International Symposium
on Personal, Indoor and Mobile Radio Communications, (PIMRC), pp. 1–5. IEEE
(2008)
16. Flammini, M., Moscardelli, L., Navarra, A., Pérennes, S.: Asymptotically optimal
solutions for small world graphs. In: Proceedings of the 19th International Confer-
ence on Distributed Computing (DISC). Lecture Notes in Computer Science, vol.
3724, pp. 414–428. Springer (2005)
17. Friedman, R., Kogan, A., Krivolapov, Y.: On power and throughput tradeoffs
of WiFi and Bluetooth in smartphones. In: Proceedings of the 30th International
Conference on Computer Communications (INFOCOM), pp. 900–908. IEEE (2011)
18. Gavoille, C., Klasing, R., Kosowski, A., Kuszner, L., Navarra, A.: On the com-
plexity of distributed graph coloring with local minimality constraints. Networks
54(1), 12–19 (2009)
19. Klasing, R., Kosowski, A., Navarra, A.: Cost minimization in wireless networks
with a bounded and unbounded number of interfaces. Networks 53(3), 266–275
(2009)
20. Korenblit, M., Levit, V.: On algebraic expressions of series-parallel and Fibonacci
graphs. In: Calude, C.S., Dinneen, M.J., Vajnovszki, V. (eds.) Discrete Mathemat-
ics and Theoretical Computer Science, pp. 215–224. Springer, Heidelberg (2003)
21. Kosowski, A., Navarra, A., Pajak, D., Pinotti, C.: Maximum matching in multi-
interface networks. Theor. Comput. Sci. 507, 52–60 (2013)
22. Kosowski, A., Navarra, A., Pinotti, M.: Exploiting multi-interface networks: con-
nectivity and cheapest paths. Wirel. Netw. 16(4), 1063–1073 (2010)
744 A. Aloisio et al.
23. Perucci, A., Autili, M., Tivoli, M., Aloisio, A., Inverardi, P.: Distributed com-
position of highly-collaborative services and sensors in tactical domains. In: Pro-
ceedings of the 7th International Conference in Software Engineering for Defence
Applications, (SEDA), LNCS. Springer (2018, to appear)
24. Valdes, J., Tarjan, R., Lawler, E.: The recognition of series parallel digraphs.
In: Proceedings of the 11th Annual ACM Symposium on Theory of Computing
(STOC), pp. 1–12. ACM (1979)
Energy Efficient Light Routing in Utility
Network
1 Introduction
Every nation has billions of building that have gas, water and electricity which
need to be measured periodically. Smart metering allows the reading of resource
consumption automatically. This technology reduces the need for manual read-
ing, decreases the annual management cost and provides real-time readings.
These help the users and the provider to reduce consumption and improve the
quality of service [3]. The implementation of smart metering systems is also
required by the Directive 2009/72/EC of the European Parliament in order to
favour the engagement of the consumers in the electricity and gas supply mar-
kets. The metering system components are devices and network. Devices also
called metric meters, perform the measurement of a resource. Their electronic
part is principally made up of simple microcontrollers with sensors. These devices
have limited capability in terms of computation, memory and power usage. For
this reason, the highest cost is not the device itself but the subsystem of inter-
connection and communication. Optical fiber, WiFI, wireless GSM and others
are amongst them. The network represents the critical part in the deployment
of the reading service [2].
Although smart meter is a branch of smart grids, the networks are extremely
different. The standard for smart grid [1] are not adequate for networks of smart
meters. For instance, the OSGP [Open Smart Grid Protocol] and ISO/IEC 14908
[1] standards have many features that are inadequate or not useful for smart
meter systems. These use Meter bus (a European standard) that includes the
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 745–754, 2019.
https://doi.org/10.1007/978-3-030-15035-8_72
746 R. Culmone and F. Pagnotta
physical [5], the link [5] (EN 13757-2) and the application layers [6] (EN 13757-
3). In particular, EN 13757-4 describes an M-bus wireless solution [4]. Although
is optional, the network layer identifies only nodes inside the network and does
not standardise any routing algorithm. Such protocol does not restrain the net-
work topology. Only the ring topology should be avoided. M-BUS is based on
a master/slave model where the supplier has the role of master and asks peri-
odically to all nodes the resource consumption. Slaves receive the request and
provide back the measured value. In most of the case, master and slave are not
directly connected but their communication is achieved via a router or repeater.
The maximum distance between a slave and a repeater is 350 m, and the max-
imum number of slaves can reach 250 [5]. This introduces the need of routing
amongst nodes. The meter communication infrastructure is frequently subject
to interference and is composed of nodes with limited memory and processing
capability. These limitations make the development of routing protocols a chal-
lenging task. Various routing approaches have been proposed in literature. In
[10,11] the authors propose an energy aware routing that is implemented as an
extension of the EN13757 Wireless M-Bus Q-mode. A cost efficient integrated
energy harvesting system powered by the available water flow was developed,
to enable operation independently of the mains grid, and eliminate the need for
battery replacement for near-zero maintenance costs.
In this paper we present Energy Efficient Light Routing in Utility Network
(EELRUN). This is a novel protocol that implements the routing without send-
ing extra control routing messages. This information (if needed) is combined
with ordinary resource information message. We have validate our protocol by
considering a network with white Gaussian noise and binary symmetric chan-
nels. The simulation shows a reduced fault while keeping a good throughput and
low power consumption.
2 Network
In our approach, each node has an error correction code table. The error correc-
tion code table of a node n assigns to each neighbour nf a value between zero
and one. This defines the error rate in the communication between nf and n.
The master node of the M-BUS has the error correction code tables of all nodes.
When a node modifies its error correction code table the new table is attached to
a resource information message in order to update the M-BUS controller tables.
In other words our protocol do not add extra control protocol messages to update
error correction code tables.
The M-BUS controller error code tables are initialised at the bootstrap phase
of the network or whenever a new slave is added. A slave announces its presence
using a hello packet. The reception of this packet triggers the initialisation of
the error correction code table. Figure 3a shows in detail the master behaviour
whenever a communication starts. The master sends a request packet in order
to receive the resource measured by a node. The packet structure is described in
Fig. 1 and contains the path to reach the destination node. The path is calculated
Energy Efficient Light Routing in Utility Network 747
by using Dijkstra algorithm where the edge connection between nodes is labelled
with the error rate. The master waits until a timeout is triggered or a response
is received.
Figure 3b shows the behaviour of a slave whenever a packet is received. A
slave can assume three kinds of roles: a listener, an intermediate or a destination
node. The former type listens to all packet in its neighbourhood and stores the
error correction code table using the Hamming error code (see Fig. 2). All the
nodes can play the listener role. The other roles instead are decided by the master
node. The intermediate node acts as a request forwarder in order to reach the
destination and whenever an intermediate node timeout occurs or a response
is forwarded back to the master together with the node correction table. The
destination node responds to the request packet using the received path and
provide back the measured resource value.
In a hop to hop communication, the path and the error correction code
table determines the dimension of the packet. When the packet size exceeds
the protocol limit, the usage of fragmentation is required. For simplicity in the
simulations, we consider a dynamic packet with no size restriction.
A packet is divided into a set of blocks. For each block, the neighbours
that retrieves the packet, computes the error correction code. Since we have no
physical layer assumption, the error correction code corresponds to the Hamming
error correction code and concerns three possible outcomes: received, damaged
but still recoverable and unrecoverable frame. The packet sender detects an
unrecoverable frame waiting the response from the next hop for a fixed amount
of time. The amount of time depends on the number of hops left to reach the
destination. This timeout feature is applied for request packets. Whenever a
timeout is triggered, the nodes of the routing path that can reach the master
provides back their error correction code table (if a variation occurs).
⎧
⎪
⎨0 ∀b ∈ B, h(b) = 0
ecc(B) = inf ∃b ∈ B, h(b) > 1 (1)
⎪
⎩ |{b∈B|h(b)=1}| otherwise
|B|
In the Eq. 1, we define the error correction code for a single packet as a function
of the set of blocks B using the Hamming function h. This returns the number
of bits in error. The cardinality of the codomain improves proportionally the
decision capability of the master node and as consequence the precision of the
master computed path.
+hamming() : Integer
+pathSize() : Integer
+removeHop(Integer)
Validate Packet
p:Packet
<<singleton>>
Choose a recipient
req : Request
Compute the shortest
path
req.removeHop(self) req.removeHop(self)
Receive
response req.pathSize() * 2k
[req.target = self]
Listener Reached
[resp.destination <> self] node [Unintelligible packet]
createRes
[Readable packet]
[resp.destination = self]
Add the node error
correction code table Hop response
Unintelligible packet in childRes timeout
[resp.success = false] [resp.success = true]
Send
childRes
Target not reached Communication Success
Hop response
2.1 Cache
In a master/slave communication, the message size affects the power consump-
tion of the node and also the number of unrecoverable messages. For each com-
munication, the master forwards a packet composed of header and payload. In
a smart meter system, the header part is bigger than the payload which is com-
posed of a routing path and a list of error correction code tables, whereas the
packet is a response type. In order to reduce the header part, we introduce a
cache system for optimise the routing path. In most of the distributed system,
the usage of caches has been deployed for different purposes [9]. This can also
be applied to routing. In our protocol, each slave has a cache table. The cache
table of a slave C assigns to a node cn the next hop. The master node has the
cache table of all nodes. As we can see from the Algorithm 2, the master do not
disseminates the routing path whereas all nodes know their next hop. On the
contrary the master includes the full path in the packet.
As we can see from the Algorithm 3, the master node introduces entries in the
cache when a communication returns. On the contrary, the Algorithm 1 shows the
removal of the cached path whether a fault occurs or a communication timeout.
Figure 4a shows the full master behaviour. Each slave updates their cache when
receives a packet with a path. Whenever the path is not provided, the node uses
the cache to get the next hop. Figure 4b shows the slave behaviour using cache.
3 Simulation
Seeing the complexity of the system, two simulations were executed for checking
the effectiveness of this protocol. In the former, the master provides the full path
whereas in the latter instead reduces the path through a distributed cache. Both
simulations require the following input parameters: messages - the number of
packets, nodes - the number of nodes in the network, connectivity - the number
of nodes reachable from each node, noise frequency - the number of packets
required for updating the noise, impact noise in the network- the percentage
of nodes affected by noise and noise edges impact - for each affected node, the
percentage of edges affected by noise. In simulations, the structure of the network
is randomly generated. The master previously chooses a list of unique slaves
and afterwards ask them cyclically the measurement. Due to having a reliable
simulation, a fundamental parameter is the number of messages. Although a high
number of packets are desirable, the simulation is not feasible for computation
reason.
In order to compute the Hamming correction code, we modelled the noise in
the network. For doing that, we introduced a weighted graph structure, described
by nodes and edges in which the weight of an edge is the noise level. During simu-
lations we alter the noise level in the network, considering a percentage of nodes.
For every node selected, the percentage of edges describes the number of edges
with a modified noise. These choices are made following a uniform distribution.
Moreover for computing the Hamming error, we consider the network as a binary
symmetric channel. Using such assumption, the Eq. 2 describe the probability of
750 R. Culmone and F. Pagnotta
<<singleton>>
Choose a
destination
req : Request
Reduce path
using cache [Unintelligible packet]
req.removeHop(self)
<next, req.target>
inside the cache.
Forward
req
req.removeHop(self)
Receive
response req.getHopCount() * 2k
newReq : Request
target = req.target
hops = req.path
source = self
Clear cache destination = next
resp : Response for hopCount = req.hopCount - 1
destination
Hop response
Target not reached Communication Success
a recipient, using the noise level r and the message size n, to receive a readable,
a damaged but recoverable and an unrecoverable packet [8]. The noise level r is
a small value between 0 and 0,5. As a noise function, we used a white Gaussian
noise (with mu = 0.25 and sigma = 0.04) for describing a real communication
between nodes.
Using the noise level and the Eq. 2, we deduce the communication outcome
through the Hamming code applied to each block. The size of a block is 16 bit
where 11 bit are used for the payload and 5 bit are for Hamming redundancy.
The simulation output without cache optimisation is about the success rate
and the throughput average of the network. The success rate is the ratio between
the number of received messages and the number of messages sent by the master.
The definition of throughput average tha is the average of the throughput in a
single hop to hop communication. This parameter represents the number of bit
sent in one tick (the smallest time unit in our simulation) with a successful
communication.
100
Successrate
50
Low
Medium 10
8
6
Noise 4
High2 %Edges
100 100
Successrate
Successrate
50 50
Low Low
Medium 10 Medium 10
8 8
6 6
Noise 4 Noise 4
High2 %Edges High2 %Edges
Fig. 5. Success rate trend (100 nodes in orange, 500 nodes in blue and 1000 in red)
Energy Efficient Light Routing in Utility Network 753
90
100
T hroughputindex
85
80
80
Low
10
8 75
Medium 6
4
Noise High2 %Edges
Fig. 6. Throughput index ti trend (red - 100 nodes, orange - 500 nodes, yellow - 1000
nodes)
For the cache simulation, we analysed only its throughput average thac. In
order to underline the cache impact, we define the throughput index ti as the
percentage of the ratio between the thac and tha.
Figure 5 shows the relation between parameters and the success rate. Such
correlations are not trivial and the simulations are a valid tool for evaluate them.
Figure 5 underlines the inverse proportional relationship between the success
rate and the noise impact. The protocol provides excellent success rate whether
the noise impact level is lower or equal to the medium one. Generally, the most
impact attribute is the network size together with the noise and the number of
edges. The communication path is influenced by all the factors. The number of
edges is not effective for lower noise level rather can be beneficial for the best
path. With greater level of noise, the edges affect more the network success rate,
especially considering bigger networks.
Figure 6 underlines the relationship between the network parameters and the
throughput index ti. As we can see, the network size holds an important role in
the throughput ti trend, which is also influenced by the number of edges. In most
of the case, the protocol does not provide any advantages using highly intercon-
nected networks. The reason concerns the minimum impact of the cache protocol
with short routing path. The cache system works well in a small network.
5 Conclusion
Our approach extends the Meter bus protocol with a noise resilient routing
protocol. This sends the path together with the message to be delivered. This
can cause a big overhead especially when the path is long. For this reason we
make use of a caching system. Our protocol performs well for medium and lower
noise level in a large networks. As future work, we plan to make a detailed energy
evaluation by using the model that is described [12].
754 R. Culmone and F. Pagnotta
References
1. Afzaal, A., Saleemi, M., Nazir, M., Saleemi, F.: Summarization of energy-efficient
communication protocols for smart grid. Int. J. Res. Rev. Comput. Sci. 169, 321–
354 (2018)
2. Saputro, N., Akkaya, K., Uludag, S.: A survey of routing protocols for smart grid
communications. Comput. Netw. 56, 11 (2012). https://doi.org/10.1016/j.comnet.
2012.03.027
3. Vannucchi, C., Diamanti, M., Mazzante, G., Cacciagrano, D.R., Corradini, F.,
Culmone, R., Gorogiannis, N., Mostarda, L., Raimondi, F.: vIRONy: a tool for
analysis and verification of ECA rules in intelligent environments. IEEE (2017)
4. Mohan, V.: An Introduction to Wireless M-Bus, Silicon Labs (2015)
5. DIN EN 13757-2: Communication systems for and remote reading of meters, Part
2: Physical and link layer. European Committee for Standardization (2005)
6. Communication systems for and remote reading of meters, Part 3: Dedicated appli-
cation layer. European Committee for Standardization (2005)
7. Data Commn and Networks (Isrd). Tata McGraw-Hill (2006). https://books.
google.it/books?id=6rRVfHq3hXoC
8. Bazlov, Y.: Coding Theory. Part 2: Hamming Distance (2010). www.maths.
manchester.ac.uk/∼pas/code/notes/part2.pdf
9. Paul, S., Fei, Z.: Distributed caching with centralized control. Comput. Commun.
(2010). https://doi.org/10.1016/S0140-3664(00)00322-4
10. Sikora, A., Werner, R., Grafmüller, J.O.: Design and implementation of an energy
aware routing extension for energy autarkic wireless M-Bus networks. In: Interna-
tional Conference on Advances in Computing, Communications and Informatics
(ICACCI) (2013)
11. Sikora, A., Werner, R., Grafmüller, J.O.: Design, implementation, and verification
of an energy autarkic, RF-based water meter with energy aware routing. In: Energy
self-sufficient Sensors; 7th GMM-Workshop (2014)
12. Ullah, Z., Mostarda, L., Gagliardi, R., Cacciagrano, D., Cacciagrano, D.: A Com-
parison of HEED based clustering algorithms - introducing ER-HEED. In: IEEE
AINA 2016, Crans-Montana, 23–25 March 2016
The 4th International Workshop on
Innovative Technologies in Informatics
and Networking (WITIN-2019)
Performance of the 25 Gbps/100 Gbps
Fullmesh RoCE Network Using Mellanox
ConnetX-4 Lx Adapter and Ruijie S6500
Ethernet Switch
Hui Li(&), Xiaolong Chen, Tao Song, Haiyin Chen, and Hao Chen
Testing Group of the Data Center Network, Ruijie Network, Fuzhou, China
{huili,chenxiaolong,songtao,chy,chenh}@ruijie.com.cn
1 Introduction
Computation, storage, and the web are the three main types of applications deployed in
data center or cloud. All these applications require IPC utilities to exchange interme-
diate data or control messages to process parallel tasks that cross over many servers. As
such, the overhead of moving data within the data center cannot be overlooked. Remote
direct memory access (RDMA) provides the ability of a device to access remote host
memory directly, without the intervention of the computer’s operating system.
The RDMA enables high-throughput, low-latency networking with low CPU utiliza-
tion, which is especially useful in parallel compute clusters.
The RDMA has the following major advantages: (1) Zero-copy, applications can
perform data transfer without the network software stack involvement and the data are
sent and received directly to the buffers without being copied between the network
layers. (2) Kernel bypass, applications that can perform data transfer directly from the
user space without the need to perform context switches. (3) No CPU involvement,
applications can access remote memory without consuming any CPU in the remote
machine. The caches in the remote CPU(s) will not be filled with the accessed memory
content.
There are several network protocols that support RDMA: (1) InfiniBand, a new
generation network protocol that supports RDMA natively from the beginning.
(2) RoCE, A network protocol that allows RDMA to be performed over Ethernet
networks. (3) iWARP, A network protocol that allows performing RDMA over
TCP. In this paper, we are deploying Remote Direct Memory Access (RDMA) tech-
nology in fullmesh RoCE network to provide ultra-low latency and high throughput
network to applications, with very low CPU overhead.
star. Figure 2 depicts a fullmesh with 11 nodes, which is a fully connected graph. In
Fig. 4, the nodes were connected to the switch in a bus topology. We also consider
Fig. 4 as a fullmesh network in logically. In Fig. 4, any two nodes are connected
through only 1 hopper or 2 hopper switch devices. As a result, the bandwidth and the
latency between any two nodes of a fullmesh network are optimized.
This paper uses OFED, MPI, and Spark to evaluate the performance of
25 Gbps/100 Gbps fullmesh RoCE network. OFED refers to OpenFabrics Enterprise
Distribution. It is a package for Linux that includes all of the needed software (libraries,
utilities and drivers) to work with RDMA enabled protocols (Infiniband, RoCE, and
iWARP). It includes a PerfTest library that we will use to evaluate the performance of
25 Gpbs/100 Gpbs fullmesh RoCE network. The message passing interface (MPI) is a
standardized and portable message-passing standard designed by a group of researchers
from academia and industry. There are several MPI runtime implementations, which
include Intel MPI, MPICH, and OpenMPI. All three MPI implementations support the
RDMA protocol. We used the OpenMPI MPI_Alltoall API to evaluate the performance
of 25 Gbps/100 Gpbs fullmesh RoCE network. Spark is a unified analytics engine for
large-scale data processing. Spark powers a stack of libraries including SQL, Data-
Frames, and MLlib for machine learning. Mellanox sponsors the SparkRDMA project
which have Spark to use RDMA protocol with only several extra configuration work of
Spark. In this paper, we also use the SparkRDMA Terasort application to evaluate the
performance of 25 Gpbs/100 Gpbs fullmesh RoCE network.
The rest of the paper is organized as follows. We provide a brief overview of the
related work in Sect. 2. We illustrate the deployment architecture of the fullmesh
network in Sect. 3. In Sect. 4, we introduce four applications and evaluate the per-
formance of 25 Gbps/100 Gpbs fullmesh RoCE network. Finally, we conclude the
paper in Sect. 5.
760 H. Li et al.
2 Related Work
2.1 Next Generation Fullmesh Network
Paper [1] evaluated the performance of the RDMA over RoCE standard in an enterprise
data centers infrastructure. It also analyzed why the RoCE has received broad industry
support from many hardware, software and system vendors. The authors in paper [2]
discussed the latest development in RoCE congestion control. Hardware congestion
control reduces the latency of the congestion control loop; the loop reacts promptly to
congestion by throttling the transmission rate quickly and accurately. The authors also
surveyed architectural features that allow deployment of RoCE over lossy networks
and presented test results. Paper [3] designed a DSCP-based priority flow control
(PFC) mechanism to ensure large-scale deployment of RoCEv2. They addressed the
safety challenges brought by PFC-induced deadlock, RDMA transport livelock, and the
NIC PFC pause frame storm problem. Their experiments showed that the safety and
scalability issues of running RoCEv2 at scale can be addressed. Paper [4] describes the
challenges of using RDMA over commodity Ethernet (RoCEv2) to support some of
Microsoft’s highly-reliable, latency-sensitive services. Their experiences show that the
safety and scalability issues of running RoCEv2 at scale can all be addressed, and
RMDA can replace TCP for intra data center communications and achieve low latency,
low CPU overhead and high throughput.
Intelligent stack [9] (iStack) connects multiple switch devices into one switch device.
iStack is a nonstandard technology in DCN that is used to extend the ports of a switch.
Since it can simplify network management, iStack has been widely used in today’s
DCN network deployment. However, iStack requires that multiple switch devices must
come from the same vendor, and be installed with the same software, which conflicts
with the openness of the next generation DCN network. Nevertheless, it is easy to
dismantle the iStack if there is a fault between switch devices. The fault of one switch
easily affects the work of other switches.
VSU lite [10] technology can overcome the disadvantages of iStack while still
providing the advantages of iStack. VSU lite utilizes dual ports to uplink two switch
devices, and NIC is set as bond mode 4, which obeys the IEEE 802.3ad (RoCE LAG).
To make use of VSU lite, it requires modification of the network driver in the Linux
kernel in addition to some configurations in the switches. Figure 4 illustrates the
762 H. Li et al.
For the best performance, we configure the PFC, ECN, DSCP, DCQCN services in
network adapters and switches for the RoCE network. Explicit congestion notification
(ECN) was initially defined for TCP/IP by embedding an indication for congestion in
the IP header and an acknowledgement in the TCP header. ECN compatible switches
and routers mark packets when congestion is detected. The congestion indication in IP
header is also used by congestion control of RoCEv2. Differentiated services code
point (DSCP) is an 8-bit field in the IP header. The network element will take the
priority from the DSCP bits and map it to the right buffer/queue/priority. DC-QCN is a
congestion control protocol for RoCE. It is based on the ECN feature on network
switches, to inform the sender to throttle the injection rate upon congestion.
Our testbed has16 Dell R720 servers. Each compute node has a 16-core 2.4 GHz
Intel Xeon E5-2630 processor with 16 GB of main memory, and it is equipped with
Mellanox ConnectX-4 LX (dual port 25 GbE/50 GbE) HCAs. We use CentOS Linux
7.4.1160 (Core) with kernel 3.10.0-693.2.2.el7.x86_64 with VSU lite patch. In addi-
tion, we use the Mellanox Open Fabrics Enterprise Distribution (OFED) [11]
MLNX_OFED_LINUX-4.2-1.0.0.
Performance of the 25 Gbps/100 Gbps Fullmesh RoCE Network 763
We ran the Terasort code on 16 nodes and sorted the 100 million records (10 GB)
within 5 min. Figure 8 show the total throughput of fullmesh network using Spark
Terasort with up to 16 nodes. The throughput is shown to be stable. It also shows the
linear scalability of the total throughput using up to 16 server nodes. However, we see
slightly fluctuations of throughput when number of nodes are 2, 3, 9, and 13. This is
because the input data is a little skewed so we could see the unfair throughput. The
result of the average throughput of each node is nearly 300 Mbps; this value is very
small because the disk I/O operations are involved in the shuffle stage of the Spark
Terasort task. In fact, even considering the peak performance of throughput of Terasort
Spark task, which is about 2 Gbps, and the total peak throughput is 2 * 16 = 32 Gpbs,
which is still much lower than the link rate of the switch.
This paper depicts the deployment architecture of 25 Gpbs/100 Gpbs fullmesh RoCE
network using VSU-lite. This paper studies the performance of throughput of
25 Gpbs/100 Gpbs fullmesh RoCE network using four applications, the MPI PingPong
jobs, the OFED scripts, MPI Alltoall primitives, and SparkRDMA Terasort. The results
show that the capability of switch and performance of disk I/O could be a performance
bottleneck in the fullmesh network. This paper also shows linear scalability of total
throughput of fullmesh network using up to 16 nodes as long as the total throughput
does not exceed the link rate of switch. We believe that our deployment architecture
and configurations of 25 Gpbs/100 Gpbs fullmesh RoCE network are optimal and can
be widely used in academia and industrial environment.
References
1. Beck, M., Kagan, M.: Performance evaluation of the RDMA over Ethernet (RoCE) standard
in enterprise data centers infrastructure. In: Proceedings of DC-CaVES (2011)
2. Guo, C., Wu, H., Deng, Z., Soni, G., Ye, J., Padhye, J., Lipshteyn, M.: RDMA over
commodity ethernet at scale. In: SIGCOMM 2016 (2016)
3. Shpiner, A., Zahavi, E., Dahley, O., Barnea, A., Damsker, R., Yekelis, G., Zus, M., Kuta, E.,
Baram, D.: RoCE Rocks without PFC: Detailed Evaluation
4. Zhu, Y., et al.: Congestion control for large-scale RDMA deployments. In:
ACM SIGCOMM (2015)
5. Islam, N.S., Rahman, M.W., Jose, J., Rajachandrasekar, R., Wang, H., Subgramoni, H.,
Murthy, C., Panda, D.K.: High-performance RDMA-based design of Hadoop mapreduce
over Infiniband. In: Proceedings of the International Conference on High Performance
Computing, Networking, Storage and Analysis, SC 2012 (2012)
6. Lu, X., Islam, N.S., Wasi-Ur-Rahman, M., Jose, J., Subramoni, H.: High-performance
design of Hadoop RPC with RDMA over InfiniBand. In: International Conference on
Parallel Processing (2013)
7. Accelerating Spark with RDMA for Big Data Processing: Early Experiences
8. Shankar, D., Lu, X., Jose, J., Wasi-ur-Rahman, M., Islam, N., Panda, D.K.: Can RDMA
benefit online data processing workloads on memcached and MySQL. ISPASS, Philadel-
phia, PA, USA (2015)
9. Stack Configuration – Huawei Technical Support. http://support.huawei.com/enterprise/
docinforeader!loadDocument1.action?contentId=DOC1000074401&partNo=10042
10. How to implement the VSU-lite in DCN network. http://www.ruijie.com.cn/fa/xw-hlw/
61546
11. https://www.mellanox.com/related-docs/prod_software/Mellanox_OFED_Linux_User_
Manual_v4.0.pdf
12. Mellanox OFED for Linux User Manual
13. https://software.intel.com/en-us/articles/intel-mpi-benchmarks
14. https://www.open-mpi.org/doc/v1.10/man3/MPI_Alltoall.3.php
15. https://github.com/ehiggs/spark-terasort
16. https://github.com/Mellanox/SparkRDMA
The Warning System for Speed Cameras
on the Road by Deep Learning
1 Introduction
According to the statistics of 2002–2016 from the Taipei City Government Police
Department, the traffic accidents still accounted for the sixth place in the top ten causes
of death, and the percentage of traffic accidents was the highest. The causes of traffic
accidents include speeding, drunk-driving, zigzag-driving and driving with negligence.
However, in fact, when driving, due to advances in vehicle manufacturing tech-
nology, the driver may unconsciously speed up. It may also be because the driver is not
familiar with the road or terrain, leading to speeding unconsciously when going
downhill. Due to these violations, drivers may get tickets and are more likely to have
car accidents. Though the speed limit guidelines for roads are clearly stated in traffic
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 768–775, 2019.
https://doi.org/10.1007/978-3-030-15035-8_74
The Warning System for Speed Cameras on the Road by Deep Learning 769
regulations, in addition to driving on highways or roads with specific speed limits, the
speed limits on ordinary roads often differ in different sections. The information of
different road sections often make drivers confused with the legal speed limit It is very
dangerous as drivers hit the brake when they are watching a speed camera or a police
car. This paper provides image capture by setting up a camera in front the car, and
provide the information of the speed limit sign to the driver through image classifi-
cation and recognition.
When the “police patrol car”, “warning sign for front speed camera”, “warning
slogan for front speed camera”, “front view of radar speed camera”, “back view of
radar speed camera”, “traffic sign of speed limit” were detected, the system would issue
a warning, and when “the traffic sign of speed limit” was detected, the number in the
middle part would be identified to obtain and update the speed limit.
Wu et al. [1] integrated the Haar-like algorithm to filter out possible detection
targets, and then used the CNN classifier to calculate the possible classification cate-
gory scores of the target, and then determined whether it was the probability of the
target and indicated the position. Yavas et al. [2] installed RFID tags in a wide range on
the traffic infrastructure, and installed RFID readers on the vehicles. When the vehicles
passed the RFID tag, they read pre-existing speed limit data in the tag.
Moreover, it has been a hot research topic in recent years on image recognition and
classification through Convolutional Neural Networks (CNN). Luo et al. [3] used the
multi-task convolutional neural network to identify traffic signs. Maurice Peemen et al.
[4] used the operation of convolutional neural networks to extract features to classify
visual items. For applications in portable and wearable devices and using hardware
accelerators to achieve and overcome the burden of computing, it could be used to
identify road speed limits on the road.
2 Design of System
As mentioned above, the paper used the embedded system to realize the warning
system for speed cameras on the road through image capture and deep learning tech-
nology. The system analyzed and identified speed limit traffic signs in real time through
image recognition technology and recorded them. While driving, “the police patrol car
on the roadside”, “the radar speed camera” and “the warning sign of front speed
camera” were recognized, the system would automatically release the warning message
through the voice and inform the maximum speed of the road.
In order to achieve the function, the hardware of the system mainly adopted
NVIDIA Jetson TX2, the CPU adopted dual-core Denver 2 64-bit CPU quad-core
ARM® A57 Complex, GPU adopted NVIDIA PascalTM architecture and 8 GB L128
bit DDR4 memory. Besides, the messages from the Bluetooth OBD-II decoder were
received via the USB Bluetooth dongle, mainly the speed of the vehicle. The images in
front of the vehicle were connected to the system via USB via SJCAM’s SJ4000 sport
camera. The diagram is shown in Fig. 1.
The development of the system was divided into training host and the embedded
vehicle implementation platform. The main training work required a lot of computing
power, therefore it had been trained on the host computer, the weight file of the trained
770 C.-C. Wu et al.
Next, when the training had been completed, the original file of darknet would be
directly modified. When judging that the output category was “traffic sign of speed
limit”, the image of “traffic sign of speed limit” would be transmitted to the digital
recognition program which was added to the system for digital identification. The
program only took hundreds and units digits for digital recognition, and finally mul-
tiplied the identification result by 10.
To recognize the numerical value of the highest speed limit correctly is very
important in this system. However, the same speed limit has different meanings in
different lanes. For example, when seeing a speed limit sign of 50 on the highway, if
the vehicle is driving on the main road of the highway, this speed limit sign cannot be
misidentified by the system unless it is to leave the interchange, as shown in Fig. 4(a)
(white circle frame). Conversely, if the vehicle is leaving the interchange to exit the
highway, the 50 speed limit sign would be meaningful, as shown in Fig. 4(b). In order
to avoid this problem, the system deliberately avoided relatively small samples when
collecting training samples. At the same time, in order to avoid recognition errors when
identifying the sign of speed limit as well as avoid misjudging the speed limit of
different lanes, the detected category would have to be in 30 30 size for digital
recognition. The system would ignore it when the detected traffic sign of speed limit
was less than the size.
Next, the weights files of Convolutional Neural Networks trained by the host had to
be ported to the embedded system. This system used NVIDIA Jetson TX2 to imple-
ment the host, the system was installed Ubuntu 16.04, and darknet had to download
source code [8]. The Digital recognition program had to be recompiled through the
development tools [9], and the trained weights file had to be copied to the NVIDIA
Jetson TX2 system. The installation architecture is shown in Fig. 5.
772 C.-C. Wu et al.
Fig. 4. The speed limit sign of leaving the interchange, (a) the screen on the main lane, and
(b) the screen when leaving the interchange
3 The Results
The parameter for training convolutional neural networks was to set 64 images of
“batch” to find features. The maximum number of iterations was set to 40,000 times.
The initial learning rate of convolutional neural network was 0.001, and after the
20,000th Batch, the learning rate was reduced 0.1 times. The final learning rate was
0.00001.
Besides, the system was mounted in the lab, as shown in Fig. 6(a). The SJ4000
sport camera was connected via USB interface to provide real-time image input of the
system. The lens position could be controlled in the laboratory through a two-axis
steering platform and connected to the Bluetooth OBD-II via USB Bluetooth Dongle
and Buzzer via GPIO. The photo of the system installed in the car is shown in Fig. 6
(b). The system used a cigarette lighter interface to provide a voltage regulator circuit
with a voltage of 12 V, which then provided battery charging and provided power to
the entire system. In the lower right corner is the photo editing of the Bluetooth OBD-II
installed under the driver’s seat. There was not enough space in the front seat, so the
two-axis steering platform was removed.
This paper is mainly to provide the speed limit for the road section while driving. It
can also provide warnings when it detects police patrol cars, radar speed cameras and
warning signs. In addition to avoiding receiving the ticket, it can avoid overspeeding on
dangerous roads. However, there was no LCD display in NVIDIA Jetson TX2, the
The Warning System for Speed Cameras on the Road by Deep Learning 773
Fig. 6 (a) Photo of the system mounted in the laboratory, (b) photo of the system installed
in the car
following result was captured in host computer. The detection result in the urban area is
shown in Fig. 7(a). The speed limit of 60 km/hr. and the speed camera were detected in
the screen, and the probability of displaying the identification result were 82% and 71%
respectively. Figure 7(b) shows the result of detection of “warning slogan for front
speed camera” and “warning sign for front speed camera” in the urban area, and the
probability of displaying the identification result were 76% and 68% respectively.
Fig. 7. Image of identification result, (a) the detection of the sign of speed limit and speed
camera in the urban area, (b) the result of detection of warning sign and warning slogan
X
N
mAP ¼ PðkÞ DrðkÞ ð1Þ
k¼1
774 C.-C. Wu et al.
The relationship between Avg_Recall and count is shown in formula (2). Recall
was defined as the ratio which was identified to be correct in the positive sample, while
count was the total number of inputs of positive samples.
DetectorNum
Avg Recall ¼ ð2Þ
count
There were 15 test videos collected, including 9 videos in urban areas and 6 videos
on the highway, which were self-captured. The result of the experiment on the YOLO
V3 when outputting 6 categories was that Mean AP was 81.25%, and Average Recall
was 85.5%. The result of the digital recognition for speed limit had a precision of
83.5% and a recall of 80.2%.
4 Conclusion
In this paper, the experiment was completed with NVIDIA Jetson TX2 and used
YOLO V3 as the architecture for CNN classification. The images were divided into six
categories: “police patrol car”, “warning sign for front speed camera”, “warning slogan
for front speed camera”, “front view of radar speed camera”, “back view of radar speed
camera”, “traffic sign of speed limit”. The result of the experiment on the YOLO V3
when outputting 6 categories was that Mean AP was 81.25%, and Average Recall was
85.5%. The result of the digital recognition for speed limit had a precision of 83.5%
and a recall of 80.2%.
Acknowledgments. This work was supported by the Taiwan Ministry of Science and Tech-
nology MOST 107-2221-E-218-023-MY2.
References
1. Wu, C.C., Wu, S.L.: A novel method for traffic sign detection and recognition. In:
Proceedings of the 2017 International Symposium on Novel and Sustainable Technology,
pp. B-12–B-13 (2017)
2. Yavas, B., Ozbilen, M.M., Yilmaz, Y.: Traffic speed sign recognition with RFID support. In:
Proceedings of the 2014 IEEE 8th International Conference on Application of Information
and Communication Technologies, pp. 1–4 (2014)
3. Luo, H., Yang, Y., Tong, B., Wu, F., Fan, B.: Traffic sign recognition using a multi-task
convolutional neural network. IEEE Trans. Intell. Transp. Syst. 19, 1–12 (2017)
4. Peemen, M., Shi, R., Lal, S, Juurlink, B., Mesman, B., Corporaal, H.: The neuro vector
engine: flexibility to improve convolutional net efficiency for wearable vision. In: Proceedings
of the 2016 Design, Automation & Test in Europe Conference & Exhibition, pp. 1604–1609
(2016)
5. Road Traffic Safety Portal Site in Taiwan. http://168.motc.gov.tw/News_Photo.aspx?n=
n1oOkV$gy2YXqPriGaYUcQ@@&sms=Qck7tL6R3WJmM6srNWCCKA@@
6. Tzutalin, D.: labelImg GitHub site. https://github.com/tzutalin/labelImg
The Warning System for Speed Cameras on the Road by Deep Learning 775
7. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: unified, real-time
object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 779–788 (2016)
8. Redmon, J.: Darknet GitHub site. https://github.com/pjreddie/darknet
9. Redmon, J.: Installing Darknet website. https://pjreddie.com/darknet/install/
Using Feature Selection to Improve
Performance of Three-Tier Intrusion Detection
System
Abstract. Social media services have become an essential part of daily life. Once
5G services launch in the near future, the annual network IP flow can be expected
to increase significantly. In case of security threats, network attacks will become
more various and harder to detect. The intrusion detection system (IDS) in the
network defense system is in charge of detecting malicious activities online. The
research proposed an intelligent three-tier IDS that can process high-speed net-
work flow and classify attack behaviors into nine kinds of attacks by seven
machine learning methods. Based on the operation time, the detection process can
be divided into the offline phase, which trains models by machine learning, and
the online phase, which enhances the detection rate of network attacks by a three-
tier filtering process. In the experiment, UNSW-NB15 was adopted as the dataset,
where the accuracy of intrusion detection approached 98%.
1 Introduction
With the recent surge of network attack events, intrusion detection has already become
one of the most important technologies to enhance the security of computer systems.
The intrusion detection system (IDS) was first proposed by Anderson [1] in 1980.
Denning [2] also applied expert system in IDS applications in 1987 and monitored user
behaviors to identify anomalous operations. As information technology advances,
researchers used machine learning with feature selection to develop IDS for improved
pattern recognition. Hall [3] adopted correlation-based feature selection with machine
learning to remove irrelevant, redundant and noise features. In 2002, Mukkamala et al.
[4] applied the neural network and Support Vector Machines (SVM) on IDS design.
Next, they merged important features selection, neural network, and SVM on IDS
construction in 2003 [5]. Kumar et al. [6] compared three feature selection methods
(correlation-based, information gain, and gain ratio) to effectively promote the accuracy
of the naïve Bayes classifier. They also proved that the feature selection method can
really optimize the machine learning classifier [7–9].
The most important feature of IDS is that it detects malicious network behaviors
from real network flow and automates related processes to protect the whole computer
system in time. The major IDS detection methods can be divided into three categories,
including misuse detection [10], anomaly detection [11], and hybrid detection [12, 13].
Misuse detection, a signature-based method, has higher accuracy than anomaly
detection, but cannot detect unknown attack behaviors. Both statistical analysis and
data mining are adopted to create a classification model for anomaly detection. The
advantage of anomaly detection is that it can detect unknown attacks, but its detection
accuracy is lower than misuse detection. Hybrid detection not only has the advantage of
the previous two detection methods but can also detect unknown network attacks.
Based on Cisco Visual Networking Index (VNI) [14] prediction, the global IP flow
will reach 4.8 ZB per year in 2022. In 2018, 3GPP also launched the first specification
of 5G and the data transfer rate of 5G is 20 Gb per second. Merging 5G and IoT, the
industry will change sharply. 5G technology will for sure bring people more conve-
niences in life, but information security works will also face extreme challenges. To
maintain high efficiency and accuracy in a high-speed network environment, this
research proposed an intelligent IDS to detect real-time network attacks. All operations
of the IDS can be divided into an offline phase and an online phase. The offline phase
adopts (1) machine learning methods to train known attack models and (2) the feature
selection method to optimize identification results. The online phase identifies mali-
cious behaviors in real-time in three-tier works. The first tier uses both a black list and a
white list to directly filter out either “normal” or “abnormal” connections with the
known websites. The tier operation can remove large amounts of network flow to skip
the detection that follows. In the second tier, the trained attack patterns are adopted to
determine whether the network flow is normal or abnormal. Only abnormal behaviors
will be sent to the third tier to check which kind of attack the abnormal behaviors
belong to. Finally, all intrusion detection results will be showed on a GUI to indicate
the current health status of the network environment.
2 Related Works
prediction results. Basically, machine learning can be divided into the following
4 categories:
• Supervised Learning: The dataset is labeled to learn data patterns. Once new data is
input, the system can use these data patterns to predict the label of the input data.
This kind of learning algorithms includes SVM, Naïve Bayes, Decision Trees,
Linear Regression, Logistic Regression and so on.
• Unsupervised Learning: Without a labeling process for data training, the machine
learning method cannot judge the accuracy of classification result. In the training
process, as soon as cases are inserted into the machine, the system can automatically
find potential rules from the input dataset. This kind of machine learning methods
includes K-Means, Hierarchical classification and so on.
• Semi-Supervised Learning: The learning method is between supervised learning
and unsupervised learning. The method uses some labeled data and large amounts
of unlabeled data for the learning work. It saves the cost of manually labeling
everything and improves the accuracy of unsupervised learning.
• Reinforcement Learning: The agents select proper actions based on observation.
They evaluate feedback rewards to accumulate experiences to reach the biggest
reward from long-term actions.
2.3 UNSW-NB15
In 2015, the Australian Defense Force Academy (ADFA) of University of New South
Wales (UNSW) Australia published an intrusion detection dataset named UNSW-
NB15 (University of New South Wale Network Based 2015) [16]. The total amount of
data in UNSW-NB15 is 2,540,044 records. Each data record contains 47 number of
features and 2 labels. One label shows that the network flow behavior is either normal
or anomalous, while the other one stands for the attack types. Basically, these 47
features are divided into 5 subgroups including Flow features, Basic features, Content
features, Time features, and Additional generated features. In addition, there are 9 kinds
of attack types in the UNSW-NB15 dataset including Analysis, Backdoors, DoS,
Exploits, Fuzzers, Generic, Reconnaissance, Shellcode, and Worms [16].
3 Research Method
The research proposed a three-tier intrusion detection system for the recognition of
malicious behaviors from real network flow. To reach higher accuracy, machine
learning and feature selection were adopted to detect attack behaviors, as shown in
Fig. 1. Based on the operation time, the system could be divided into two operations,
the offline phase and the online phase.
IDS directly removed these network flows from the checking list, e.g. Facebook,
Google, YouTube and so on. The first-tier filtering process can remove lots of highly
dangerous/safe network flow and largely reduce the detection load of the second tier.
Before the detection work, the second-tier loaded 7 kinds of machine learning models
trained in the offline phase. If all trained models cannot find any attack behavior from
the current connection, the connection directly passed the third-tier checking. Once the
training models voted that the current connection was an attack, the connection would
be sent to the third-tier to identify the attack.
In the offline phase, the proposed IDS adopted machine learning methods to train
attack patterns for optimizing online phase operations. These trained models were used
to detect the network attacks. The related operations are shown in Fig. 2.
krcf
Merits ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð1Þ
k þ kðk 1Þkrff
where k represents the number of features, rcf is the average correlation value between
features, and their class, rff shows the average correction value between features.
The Correlation method computes the value of Pearson’s correlation coefficient
between features and their classes. A higher value means that the feature is more
suitable to predict the class of input data. The ranking list of the feature set, including
the preset number of features, follows the descending order of correlation values.
The Information Gain method evaluates the feature worth as shown in Eq. (2).
A higher value also means that it is good for data classification. Generally, Information
Gain contains features with different values, but the values do not signify importance,
for example source IP or destination IP.
H ðClassÞ HðClassjAttributeÞ
GRðClass; AttributeÞ ¼ ð3Þ
H ðAttributeÞ
The Wrapper method adopts learning algorithms to evaluate the feature dataset. In
the evaluation of k-folder cross-validation, each selected feature set has to be selected
by learning algorithms with results evaluated. Therefore, the process causes large
amounts of computation, but can select a feature set with higher quality.
4 Experimental Results
UNSW-NB15 [17, 18] was adopted to be the experiment dataset in this study. The
main reason is that the network behavior of the dataset is similar to that of the current
network flow. The dataset includes nine kinds of the attack classifications, which is also
closer to current attack behaviors. The large amount of data in the dataset is expected to
lead to more reliable evaluation result.
The machine learning methods [23], once adopted, converted the intrusion detec-
tion dataset to meaningful data. First, seven machine learning methods (BayesNet,
NaiveBayes, JRip, PART, J48, RandomTree and RandomForest) of Weka were
adopted to train the attack models. Each method was applied to only part of the features
selected to construct feature patterns, which were in turn used to identify nine kinds of
network attacks.
The specification of experiment environment is CPU Intel(R) Core i7-7700,
memory 8 GB, operating system Microsoft Windows 10. The Weka version is 3.8. The
experiment dataset was extracted from UNSW-NB15 where 440,044 records were
randomly selected as the training dataset and 200,000 records as the testing dataset.
Seven machine learning methods of Weka were used to train attack patterns, but they
were applied only to part of the features selected by 4 feature selection methods, as
shown in Table 1. To train each attack pattern with higher accuracy, each machine
learning method adopted a different feature selection method to select a different
number of features, e.g. 9, 12, 15. In general, the number of features ranges from
9 to 15 [24, 25].
Table 2 shows that seven machine learning methods were adopted with or without
feature selection methods to train the patterns of network attack behaviors. Most of the
time, the number of features from 9 to 15 had better results [25]. In the experiment, the
numbers of selected features were 9, 12, and 15. The measure results included recall,
precision, FNR (False Non-discovery Rate), AA (Average Accuracy), and MCC
(Matthews Correlation Coefficient). The execution time was effectively reduced when
using feature selection, and the value of AA also showed improvement.
FNR (False Non-discovery Rate): all anomalous network behaviors are erroneously
judged as normal network behaviors (the expect value is 0), as shown in Eq. (6).
FN
FNR ¼ 100 ð6Þ
TP þ FN
AA (Average Accuracy) can be computed, as shown in Eq. (7), and the expected
value of AA is 100.
TP þ TN
AA ¼ 100 ð7Þ
TP þ TN þ FP þ FN
TP TN FP FN
MCC ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8Þ
ðTP þ FPÞðTP þ FN ÞðTN þ FPÞðTN þ FN Þ
5 Conclusion
The research proposed a real time IDS that comprises an offline phase and an online
phase, to detect malicious network behaviors with behavior patterns trained by 7
machine learning methods. To improve the performance and accuracy, 4 feature
selection methods were adopted in the proposed IDS. The experiment adopted the
UNSW-NB15 dataset, which includes 9 kinds of attack behaviors, to test the accuracy
of the IDS. Experimental results show that IDS using feature selection to train the
attack models can have better results in execution time and average accuracy.
References
1. Anderson, J.P.: Computer Security Threat Monitoring and Surveillance (1980)
2. Denning, D.E.: An intrusion-detection model. IEEE Trans. Softw. Eng. 13(2), 222–232
(1987)
3. Hall, M.A.: Correlation-based feature selection for machine learning. Doctoral Dissertation,
University of Waikato, Department of Computer Science (1999)
4. Mukkamala, S., Janoski, G., Sung, A.: Intrusion detection using neural networks and support
vector machines. In: Proceedings of the International Joint Conference on Neural Networks.
vol. 2, pp. 1702–1707 (2002)
5. Sung, A.H., Mukkamala, S.: Identifying important features for intrusion detection using
support vector machines and neural networks. In: Proceedings of International Symposium
on Applications and the Internet.vol. 217, pp. 209–216 (2003)
6. Kumar, K., Batth, J.S.: Network intrusion detection with feature selection techniques using
machine-learning algorithms. Int. J. Comput. Appl. 150(12), 1–13 (2016)
7. Amiri, F., Yousefi, M.R., Lucas, C., Shakery, A., Yazdani, N.: Mutual information-based
feature selection for intrusion detection systems. J. Netw. Comput. Appl. 34(4), 1184–1199
(2011)
8. Li, Y., Xia, J., Zhang, S., Yan, J., Ai, X., Dai, K.: An efficient intrusion detection system
based on support vector machines and gradually feature removal method. Expert Syst. Appl.
39(1), 424–430 (2012)
9. Ambusaidi, M.A., He, X., Nanda, P., Tan, Z.: Building an intrusion detection system using a
filter-based feature selection algorithm. IEEE Trans. Comput. 65(10), 2986–2998 (2016)
10. Scarfone, K., Mell, P.: Guide to intrusion detection and prevention systems (IDPS). NIST
Special Publication, 800-94 (2007)
11. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection: a survey. ACM Comput. Surv.
(CSUR), 41(3) (2009)
12. Depren, O., Topallar, M., Anarim, E., Ciliz, M.K.: An intelligent intrusion detection system
(IDS) for anomaly and misuse detection in computer networks. Expert Syst. Appl. 29(4),
713–722 (2005)
13. Kim, G., Lee, S., Kim, S.: A novel hybrid intrusion detection method integrating anomaly
detection with misuse detection. Expert Syst. Appl. 41(4), 1690–1700 (2014)
14. Cisco Visual Networking Index: Forecast and Trends, 2017–2022. https://www.cisco.com/c/
en/us/solutions/collateral/service-provider/visual-networking-index-vni/white-paper-c11-
741490.html#_Toc529314175
15. Blum, A.L., Langley, P.: Selection of relevant features and examples in machine learning.
Artif. Intell. 97(1–2), 245–271 (1997)
Using Feature Selection to Improve Performance 785
16. Moustafa, N., Slay, J.: UNSW-NB15: a comprehensive data set for network intrusion
detection systems (UNSW-NB15 network data set). In: Proceedings of Military Commu-
nications and Information Systems Conference (MilCIS), pp. 1–6 (2015)
17. Moustafa, N., Slay, J.: A hybrid feature selection for network intrusion detection systems:
Central points. In: Proceedings of the 16th Australian Information Warfare Conference,
pp. 5–13 (2015)
18. Moustafa, N., Slay, J.: The significant features of the UNSW-NB15 and the KDD99 data sets
for network intrusion detection systems. In: Proceedings of 4th International Workshop on
Building Analysis Datasets and Gathering Experience Returns for Security (BADGERS),
pp. 25–31 (2015)
19. Weka 3: Data Mining Software in Java. https://www.cs.waikato.ac.nz/ml/weka/
20. Hall, M.A., Holmes, G.: Benchmarking attribute selection techniques for discrete class data
mining. IEEE Trans. Knowl. Data Eng. 15(6), 1437–1447 (2003)
21. Holte, R.C.: Very simple classification rules perform well on most commonly used datasets.
Mach. Learn. 11(1), 63–90 (1993)
22. Kononenko I.: Estimating attributes: Analysis and extensions of RELIEF, Lecture Notes in
Computer Science, vol. 784, pp. 171–182. Springer, Heidelberg (1994)
23. Prachi, C.: Usage of machine learning for intrusion detection in a network. Int. J. Comput.
Netw. Appl. 3(6), 139–147 (2016)
24. Al-Jarrah, O.Y., Siddiqui, A., Elsalamouny, M., Yoo, P.D., Muhaidat, S., Kim, K.: Machine-
learning-based feature selection techniques for large-scale network intrusion detection. IEEE
34th International Conference on Distributed Computing Systems Workshops (ICDCSW),
pp. 177–181 (2014)
25. Rathore, M.M., Ahmad, A., Paul, A.: Real time intrusion detection system for ultra-high-
speed big data environments. J. Supercomput. 72(9), 3489–3510 (2016)
A CNN-Based Method for Infant Cry
Detection and Recognition
1 Introduction
In recent years, a lot of deep learning research had been applied to image recognition
and speech recognition. Researchers had also proposed many algorithms such as
convolutional neuron networks (CNN), Long Short Term Memory Network, LSTM,
GoogLeNet, and Generative Adversarial Network [1–4]. In addition, Yamamoto had
successfully identified infant’s cry using fast Fourier transform (FFT) and PCA [5–7].
Yizhar had great success on the recognition of infant cries through the combination of
algorithms such as the Mel Frequency Cepstral Coefficient (MFCC), Spectrogram and
CNN [8–11]. Based on the aforementioned research results, it is obvious that deep
learning is an effective and suitable algorithm for the discrimination of infant cries.
Therefore, in this work, the authors propose to use pre-emphasis, spectrogram, and
CNN to identify crying, as well as a two-stage method to classify the crying into four
categories.
2 Methodology
The proposed method differentiates between infant cries caused by pain, hunger,
sleepiness, and wet diapers using deep learning with two stages. Figure 1 shows the
system architecture of the proposed two-stage system. First of all, the sampling signal
2.1 Pre-emphasis
The pre-emphasis stage refers to a system process designed to increase (within a
frequency band) the magnitude of some (usually higher) frequencies with respect to the
magnitude of other (usually lower) frequencies. Low frequency indicates slow varia-
tions in time and the numerical values of a low-frequency signal tend to change slowly
or smoothly from sample to sample. Due to the aforementioned property, Eq. (1) is
used to implement the pre-emphasis process (in this paper a ¼ 0:95) [12].
Figure 2 shows that the pre-emphasized signal becomes more pronounced in the
treble part, and the sound’s intervals become more defined.
788 C.-Y. Chang and L.-Y. Tsai
2.5 Dataset
The infant cry dataset in this paper consists of samples recorded by first-time parents
using a smart phone with a mono channel, 8 kHz sampling rate and 16-bit resolution.
The samples were stored in WAV audio format. The dataset contains a total of 19691
samples. The length of each sound file is five seconds. The dataset was collected over
nine months from the birth of the infant. The dataset is tabulated in Table 1 and the
number in each field represents the samples recorded in a month.
790 C.-Y. Chang and L.-Y. Tsai
3 Experiments
In order to maintain the balance of the training set during data training, 1500
samples randomly selected from each category of infant crying (Pain, Hunger,
Sleepiness, Wet diaper), were considered as positive crying samples. In addition,
another 1500 samples randomly selected from non-infant crying categories were
considered as negative samples.
4 Conclusion
Early stage infants can only express their emotions or physiological needs through
crying. Moreover, non-infant cry data will reduce the accuracy of infant-cry recogni-
tion. To solve the problem, this paper proposes a two-stage method to eliminate the
non-cry samples and recognize the cause of the cry. The first stage uses re-emphasis
and normalized mechanisms to convert the samples to spectrograms, which are used as
CNN model input. The proposed method achieves an accuracy of 99.83% for infant
crying detection. In the second stage, the second model divides the reasons for crying
into four types (Pain, Hunger, Sleepiness, Wet Diaper), and the recognition accuracy
reaches 78.27%.
792 C.-Y. Chang and L.-Y. Tsai
References
1. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
2. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105
(2012)
3. Ballester, P., de Araújo, R.M.: On the performance of GoogLeNet and AlexNet applied to
sketches. In: AAAI (2016)
4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville,
A., Bengio, Y.: Generative adversarial nets. In: Advances in Neural Information Processing
Systems, pp. 2672–2680 (2014)
5. Rader, C., Brenner, N.: A new principle for fast Fourier transformation. IEEE Trans. Acoust.
Speech Signal Process. 24(3), 264–266 (1976)
6. Tyagi, V., Wellekens, C.: On desensitizing the Mel-Cepstrum to spurious spectral
components for robust speech recognition. In: Proceedings of the IEEE International
Conference on Acoustics, Speech, and Signal Processing, (ICASSP 2005) , vol. 1 (2005)
7. Garcia, J.O., Garcia, C.R.: Mel-frequency cepstrum coefficients extraction from infant cry
for classification of normal and pathological cry with feed-forward neural networks. Neural
Netw. 4, 3140–3145 (2003)
8. Petroni, M., et al.: Identification of pain from infant cry vocalizations using artificial neural
networks (ANNs). In: Applications and Science of Artificial Neural Networks, vol. 2492.
International Society for Optics and Photonics (1995)
9. Yong, B.F., Ting, H.N., Ng, K.H.: Baby cry recognition using deep neural networks. In:
World Congress on Medical Physics and Biomedical Engineering 2018. Springer, Singapore
(2019)
10. Abdel-Hamid, O., et al.: Convolutional neural networks for speech recognition. IEEE/ACM
Trans. Audio Speech, Lang. Process. 22(10), 1533–1545 (2014)
11. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems (2012)
12. Muda, L., Begam, M., Elamvazuthi, I.: Voice recognition algorithms using Mel Frequency
Cepstral Coefficient (MFCC) and Dynamic Time Warping (DTW) techniques. arXiv preprint
arXiv:1003.4083 (2010)
13. Sohn, J., Sung, W.: A voice activity detector employing soft decision based noise spectrum
adaptation. Acoust. Speech Signal Process. 1, 365–368 (1998)
14. Fushiki, T.: Estimation of prediction error by using K-fold cross-validation. Stat. Comput. 21
(2), 137–146 (2011)
The Sensor Calibration and Growth
Parameters Monitoring for Phalaenopsis
Cultivation
Abstract. In this paper, a sequence of experiments that calibrate and collect the
sensor data that monitoring the growth of the orchids are derived. the sensor’s raw
data is calibrated and compared by using the popular and trusted WET2-2-K1
sensor. We also develop a cloud server and an analysis algorithm that collect the
data via different sensors by IoT technologies. A front-end responsive webpage
that visualize the Phalaenopsis physiological sensing parameters such as tem-
perature, moisture and conductivity of the cultivated soil are built, users can
monitor data on various devices such as personal computers, laptops or smart
hands to control and monitor Phalaenopsis’s growth environment data at any
time.
1 Introduction
According to the United Nations Commodity Trade Statistics Database (UN Comtrade),
orchids have a global output value of $1.7 billion in 2015 [1]. The Phalaenopsis flowers
have beautiful appearance, long flowering period and diverse colors, and have the
highest economic value in Taiwan. Taiwan is located in the subtropical zone; the good
weather conditions make Taiwan become an ideal growing environment for most
orchids. Taiwan is the northernmost boundary of the distribution of the original species
of Phalaenopsis. Over the past ten years, with the active guidance of the Taiwan’s
agricultural administrative units and the efforts of the orchid industry, the production
technology has increasingly improved. The transportation and marketing system have
gradually upgraded, and the planting area has also expanded year by year. The popu-
lation participating in cultivating orchids industry is the largest part in the flower
industry. According to the statistics of the Taiwan Customs Department (2018), the
export value of orchids in 2016 reached US$172 million, and Phalaenopsis accounts for
an export value of US$129 million.
The superior orchid planting process can ensure the quality, effectively reduce the
cost and increase the output value. It can not only promote the development of this
industry, but also increase the international competitiveness. Phalaenopsis has various
growth-related factors from seedling to flowering stage, all the stages needs to be
properly controlled and analyzed to avoid pests or diseases to affect the flower quality.
In this paper, a series of reliable experiments were deduced that calibrate and
collect sensor data that monitors orchid growth. The raw data of the sensor is calibrated
and compared using the trusted and widely used Delta-T Devices WET2-2-K1 sensor
[2]. We use IoT technology to collect different sensor data and store it in the cloud
server. Users are allowed to control and monitor Phalaenopsis physiological sensing
parameters (such as temperature, moisture and soil conductivity) which are displayed
on the front-end responsive web page via various devices (PC, laptops or mobile
devices) at any time [3].
2 Background
The advantages of Taiwan’s orchid industry are mostly focused on the production
aspects, such as variety, climate and cultivation techniques. The most suitable tem-
perature for Phalaenopsis growth is between 25–30 °C. In summer, the temperature is
mainly controlled by the inner and outer shading nets, water walls and fans in the
greenhouse; while in winter, the heating machine is used to avoid the cold damage.
There is no complete and optimized physiological model can be intelligently referenced
and universally applied to other producers, places and varieties.
Taiwan’s production technology using sphagnum moss as a single cultivation
medium is quite sophisticated. However, the density, watering, frequency and nutrient
management methods of the sphagnum moss medium used by the industry are dif-
ferent, resulting in uneven seedling quality. Electrical conductivity (EC) is one of
the important factors affecting the vegetative growth of Phalaenopsis. The EC value is
the degree of conductivity of the salt dissociated in water, and can be regarded as the
concentration of soluble salts present in the medium. The higher the concentration, the
higher the osmotic potential; making crops difficult to absorb water and nutrients. As
the result, the root is damaged or leaves are withered. How to maximize the good
characteristics of the sphagnum moss medium, and to establish optimal physiological
parameters is important to improve the quality of the seedlings.
The concept of smart agriculture is to employ sensing components such as
biosensor, environmental sensors and image recognition sensors to integrate the IT and
IoT technologies to reduce the labor demand in farming and provide a more efficient
model to cultivate crops [4]. The Phalaenopsis flowers are popular crop and owns the
high-value in Taiwan for its beautiful flower shape. The advantages of Taiwan’s orchid
industry are mostly focused on production aspects, such as flower variety, climate and
cultivation techniques. Orchids are cultivated in greenhouse, the main factor that
affecting the growth condition contains the temperature, the moisture, the illumi-
nometer, soil moisture and the carbon dioxide density.
The first usage of IoT technology in the Phalaenopsis environmental monitoring
has adopted a non-destructive method to measure the growth parameters. In 2014, Peng
et al. [5] proposed a method to measure physiological parameters by using the spec-
trometer. The NTU team proposed a greenhouse monitoring system that applies the IoT
techniques such as wireless environment monitoring, image acquisition and back-end
The Sensor Calibration and Growth Parameters Monitoring for Phalaenopsis Cultivation 795
In this study, we selected the WET2-2-K1 sensor (as shown in Fig. 1(a)) from Delta-T
Devices as the benchmark to calibrate the proposed sensors. The WET2-2-K1 is widely
used in horticulture and plant research, but this sensor lacks data upload function.
Therefore, when performing the calibration process, we must first manually measure
and record the measurement time, HubID and other data, and then upload it to the cloud
server. The newly developed moisture and conductivity sensor from Hugreen [7] are
employed for calibration purpose. Hugreen is a startup company that building precision
agriculture solutions to equip greenhouses and smallholder farmers with the tools
needed to become profitable. The proposed moisture and EC sensors are shown in
Fig. 1(b)(c). Since no sensor is perfect, the sensor must be calibrated before use for the
best accuracy. The signals of all measurement systems are affected by random noise or
signal lag. The reasons for calibration the sensors are to verify the precision (same input
796 D.-H. Chen et al.
should derive the same output) and resolution (reliably detect the small change in
measurement parameter) [8, 9]. In this study, we will perform a calibration of the
moisture and conductivity sensors. The calibration steps are as follows.
Fig. 1. The sensor used in this study: (a) WET2-2-K1 from Delta-T Devices (b) moisture sensor
from Hugreen (c) EC sensor from Hugreen
The Delta-T WET2-2-K1 has two measurement modes for measuring soil, namely
mineral and organic. Although the sphagnum moss is an organic material, the mineral
mode is often used in practical Phalaenopsis cultivation. In this study, we measure the
sensor data in two modes. The invalid samples of the sensors data (environmental tem-
perature, illuminometer, soil temperature, moisture and electrical conductivity) are first
removed, then we analyze the relation between the data from the Delta-T WET2-2-K1
sensor and the sensor to be calibrated.
4. Repeat step 3 by adding 5 g water each time until at least 100 g water is added
(the theoretical moisture is 50%).
The moisture sensor calibration experiment result:
The experimental result of the moisture sensor calibration in mineral mode is
shown in Fig. 2. We can find that the correlation coefficient between Delta-T WET2-2-
K1 sensor and the proposed sensor is 0.77 before calibration; and the value become
0.86 after calibration. Although there is a difference between two sensors, the differ-
ence remains consistent; that means the accuracy of the proposed sensor is much
similar comparing with the Delta-T WET2-2-K1 sensor.
Fig. 2. The moisture sensor calibration result in mineral mode (a) before calibration (b) after
calibration
The experimental result of the moisture sensor calibration in organic mode is shown
in Fig. 3. The correlation coefficient is 0.75 before calibration; however, the correlation
coefficient is 0.19 due to an unexpectable error condition. The left part result in Fig. 3(b)
shows that the data of the two sensors are very consistent and the experiment design is
feasible. This result leaves a problem that has yet to be resolved.
Fig. 3. The moisture sensor calibration result in organic mode (a) before calibration (b) after
calibration
798 D.-H. Chen et al.
200 0:8
V¼ ¼ 7:5
8
1. Experiment 1: Add 60 g of the 30% water percentage DFS to the sphagnum moss to
the beaker. Gradually dilute the concentration of the solution by adding water (7.5 g
water each time) as shown in Fig. 4(a).
2. Experiment 2: Add 80 g of the 40% water percentage DFS to the sphagnum moss to
the beaker. Gradually dilute the concentration of the solution by adding water (10 g
water each time) as shown in Fig. 4(b).
3. Experiment 3: Add 100 g of the 50% water percentage DFS to the sphagnum moss
to the beaker. Gradually dilute the concentration of the solution by adding water
(7.5 g water each time) as shown in Fig. 4(c).
The electrical conductivity (EC) experiment result:
As shown in Figs. 5 and 6, we perform the electrical conductivity experiment in
mineral mode and organic mode, respectively. The Ecb is the measured data and Ecs is
the converted data by Hugreen’s transformation equation. Comparing with the data by
Delta-T WET2-2-K1, the correlation coefficient before calibration is better than the
correlation coefficient after calibration. Possible causes of unsatisfactory results include
the poor design of the experimental method, errors in the measurement process, and
new sensor design issues. The results show that the proposed calibration method still
needs more experiments to verify its feasibility.
Fig. 5. The EC sensor calibration result in mineral mode (a) before calibration (b) after
calibration
Fig. 6. The EC sensor calibration result in organic mode (a) before calibration (b) after
calibration
800 D.-H. Chen et al.
The binary raw data (including soil temperature, moisture, illuminometer and con-
ductivity) returned by the Phalaenopsis sensor installed in the greenhouse (as shown in
Fig. 7) will be converted to decimal by the program. Users can periodically set up
sensors to capture data every 1 min, 10 min, 30 min or 1-h intervals. The sensor
specifications are shown in Table 1.
The data collected by the sensor is transmitted from the hub to the gateway via
LoRa (Long Range) technology and then transmitted to the cloud database via Wifi.
LoRa is a low-power wide area network (LPWAN) communication protocol that
provides inexpensive remote connectivity for IoT devices in rural, remote and offshore
industries [10]. The overall architecture is illustrated in Fig. 8. In order to let users
query the growth history of Phalaenopsis, and to develop an artificial intelligence
algorithms in the future, we have set up a cloud server to store sensor data. The system
includes a back-end Apache web server, and the front-end web page accesses the
database via PHP. The sensor data is filtered according to the date selected by the user
and encoded into a JSON format. The application programming interface (API) is
The Sensor Calibration and Growth Parameters Monitoring for Phalaenopsis Cultivation 801
provided to allow different front-end devices to display sensor data with a responsive
web page.
The sensor returns the environmental data to the cloud server, and user can access
the growth data of Phalaenopsis including user information, Hub number, Gateway
number, as well as power, environment temperature and humidity, illuminometer, soil
temperature and soil moisture of the sphagnum moss, and soil conductivity. The
growth parameters are shown in Fig. 9.
5 Conclusion
In this paper, a sequence of experiments that calibrate and collect the sensor data that
monitoring the growth of the orchids are derived. the sensor’s raw data is calibrated and
compared by using the popular and trusted WET2-2-K1 sensor. We also develop a
cloud server and an analysis algorithm that collect the data via different sensors by IoT
technologies. A front-end responsive webpage that visualize the Phalaenopsis physi-
ological sensing parameters such as temperature, moisture and conductivity of the
cultivated soil are built, users can monitor data on various devices such as personal
computers, laptops or smart hands to control and monitor Phalaenopsis’s growth
environment data at any time.
The experimental results show that the calibrated sensor data is indeed close to the
value of WET2-2-K1. However, the experimental results show that the newly devel-
oped sensor still has stability problems, which leads to sometimes the numerical dif-
ference is too large, which affects the correlation coefficient between the two sensors.
The experimental data obtained in this study will be provided to the sensor developer
for reference to further improve the stability of the sensor. At the same time, we will
experiment with different calibration methods and find other unexplained causes of
numerical fluctuations, so that the sensor can be used accurately for Phalaenopsis.
Acknowledgement. The author thanks the Ministry of Science and Technology in Taiwan for
funding (MOST 107-2321-B-415-005-) this study.
References
1. UN Comtrade (2015). http://comtrade.un.org/data/. Accessed 18 Sept 2015
2. WET KIT and WET-2 Sensor. https://www.delta-t.co.uk/product/wet-2-horticulture/
3. Gardner, B.S.: Responsive web design: enriching the user experience. Sigma J. Digit.
Ecosyst. 11(1), 13–19 (2011)
4. Dan, L., Xin, C., Chongwei, H., Liangliang, J.: Intelligent agriculture greenhouse
environment monitoring system based on IOT technology. In: 2015 International Conference
on Intelligent Transportation, Big Data and Smart City, pp. 487–490 (2015)
5. Peng, Y.-H., Hsu, C.-S., Huang, P.-C., Wu, Y.-D.: An effective wavelength utilization for
spectroscopic analysis on orchid chlorophyll measurement. In: 2014 IEEE International
Conference on Automation Science and Engineering (CASE), pp. 716–721 (2014)
6. Madakam, S., Ramaswamy, R., Tripathi, S.: Internet of Things (IoT): a literature review.
J. Comput. Commun. 3(05), 164 (2015)
7. Hugreen Co Ltd. https://hugreen.com.tw/
8. Sudduth, K.A., Drummond, S.T., Kitchen, N.R.: Accuracy issues in electromagnetic
induction sensing of soil electrical conductivity for precision agriculture. Comput. Electron.
Agric. 31(3), 239–264 (2001)
9. Quinones, H., Ruelle, P.: Operative calibration methodology of a TDR sensor for soil
moisture monitoring under irrigated crops. Subsurf. Sens. Technol. Appl. 2(1), 31–45 (2001)
10. LoRa. https://en.wikipedia.org/wiki/LoRa
The 4th International Workshop on Big
Data Processing in Online Social
Network (BOSON-2019)
Automatic Text Preprocessing
for Intelligent Dialog Agents
1 Introduction
The present research focuses on a hybrid module for the preprocessing of Italian
texts. It represents a part of the research and the industrial project titled Big
Data for Multi-Agent Specialized System (BIG4MASS), which is developed by
the Network Contacts enterprise1 in collaboration with the Department of Polit-
ical and Communication Science of the University of Salerno and co-financed by
the Ministry of Economic Development. The BIG4MASS project concerns the
development of an Hybrid and distributed Question Answering System composed
by a set of Dialog Bots able to process great volumes of data and provided with
a Syntactic/Semantic Linguistic Engine.
In the last decade the information present on the web has increased more and
more. Talking about Big Data means talking about different kind of information
which can have a strong impact on different knowledge areas such as health [3],
government [1,2] or corporate. Due to the strategic importance of unstructured
data in information retrieval and domain specific knowledge management, a
1
Network Contacts is one of the major national player in BPO services (Business Pro-
cess Outsourcing), CRM (Customer Relationship Management), Digital Interaction
and Call & Contact Center.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 805–814, 2019.
https://doi.org/10.1007/978-3-030-15035-8_78
806 A. Maisto et al.
In this Section we briefly mention the state of the art works related with the
three main modules of the preprocessing analysis.
Thanks to the increased availability of annotated corpora, it became possible
to train classifiers in order to automatically correct specific errors within texts
[16,29]. Classifiers models perform well over classes of single errors, failing when-
ever a classification system has to correct multiple (and often interconnected)
errors within a phrasal unit. To overcome this issue, hybrid methodologies have
been proposed [10]. In order to enhance sequential error analysis and to reach
better overall performances, Machine Translations algorithms have been tested
out. The logic behind automatic correction via machine translation is to con-
sider the correction itself as a translation task. Correction systems have two core
components: a linguistic model and a translation model [23]. Different kind of
models have been proposed such as hybrid approaches where a statistic algorithm
is integrated with a rule-based linguistic model [14], statistic approaches embed-
ded with the Levenshtein’s distance measure [19] and neural network approaches
[5,35,36].
Automatic Text Preprocessing for Intelligent Dialog Agents 807
As regards clitics, we define them as linguistic units that combine with other
words, called in literature “hosts”, with which they are phonologically related
[30]. In the present work, with “clitic” we mean just Italian clitic pronouns.
When clitics and verbs are agglutinated together they must be subjected to
a set of morpho-syntactic restrictions [30]. A software able to correctly analyze
and decompose clitic particles must take into account such constrains.
The present research focuses on agglutinated verbal forms, where verbs and
clitic particles appear to be tied together (with the clitic occurring after the
verb) into “monorematic” words (without any blank space between the words).
Such verbal forms, differentiated on the base of their moods, are of these kinds:
In all the other cases, which wont be object of the present research, clitics
and clitic clustering (sequences of more than one clitic disposed in succession),
occur before the verb, separated by blanks (e.g. ci si gioca).
Among the studies focused on clitics in Italian we mention [12,24,25,30,38].
As far as Pos Tagging is concerned, we must say that it is considered a
“solved task”, with state-of-the-art taggers achieving precision of 97%–98% [32,
37]. Among the best algorithms proposed in literature, we mention the Averaged
Perceptron methods [8], Stanford Tagger [22], SVMTool [15], BI-LSTM-CRF
[17], and NLP4J [6].
In Lemmatization task, tokens are converted into their common base form:
the lemma. In Italian language, the masculine singular is used as canonical form
of nouns, adjectives, pronouns and determiners, the present infinitive tense is
used for verbs. Often, Lemmatization can be replaced by Stemming, a procedure
that removes suffixes from a word form until the “stem” was found [28]. Actually,
both the approaches can be combined as in [18] for Icelandic language. Ripple
Down Rules [34] have been developed for knowledge acquisition and maintenance
of rule-based systems. The algorithm create exception to existing rules when a
rule produce incorrect conclusions. Most algorithms use external lexical informa-
tion, as Morfette [7], or SEM [9] for French texts. In particular, Morfette uses a
probabilistic system in order to joint morphological tagging and lemmatization
from morphologically annotated corpora. For what concern Italian resources, we
can cite an Italian version of TreeTagger [31], TagPro [27], CORISTagger [13],
Tanl POS tagger [4], ensemble-based taggers [11] and Pattern tagger [33].
The present section will illustrate the methodologies involved in the development
of the three different modules of the project and the results of the experiments
carried out to proof the efficiency of the systems.
808 A. Maisto et al.
• StartUp: this phase is based on the distance algorithm and its purpose is to
prepare the data structure for the style correction. Starting from the correct
dictionary entry (e.g. acqua), the system checks the maxEditDistance value:
with the value of 1 the system cyclically subtracts each character of the word,
generating a list of possible errors (e.g. aqua; acua; cqua; acqu; acqa); with a
value of 2, the algorithm generates, in addition to the previously mentioned
list, another list where the items are created by cyclically subtracting couples
of characters from the correct word (e.g. aqu; cqu; aua; cua; acq, among
others). The purpose is to create a Knowledge Base (KB), created by the
tool, that includes an extended dictionary containing the lists of errors.
• Ready: this step includes two activities:
– the correction of simple words taken in isolation: a word W is compared
with the words contained in the correct dictionary. If there is a match with
a KB word, then the tools assumes that the word W has been written
in the proper way. Otherwise, it controls in the lists of errors contained
into the KB whereas there is a string close to W (DamerauLevenshtein
distance).
– correction of words connected by apostrophes: the system considers the
expressions composed by determiner/preposition + apostrophe + noun
just as simple words, to be listed and processed into the KB.
– the correction of multiword expressions (MWE)2 : the correction algo-
rithm treats these expressions just as patterns which are analyzed as a
whole, in order to avoid the alteration of the meaning of the MWE.
The Style Correction system has been tested on an experimental set of 1500
sentences (about 15,600 tokens) - each of which contained at least one or more
typing error - selected manually from an original corpus at disposal of Network
Contacts and made up of customer answers within the TEL-CO domain. As out-
lined in Table 1, the system corrected 2141 errors out of 2362. The vast majority
of these errors regarded the orthographic forms of single unit words, a small
amount were related to misspelled entity like smartphone names or telephone
promotion denomination which are usually lexically configured as multiword
expressions. The remaining errors can be explained as syntactical errors and
were not detected by the algorithm architecture.
2
With MWE we make reference both to lemmatized MWE, e.g. carta di credito, which
are listed in the Knowledge base, and to a set of MWE which are domain specific,
e.g. all inclusive unlimited, for the domain of telecommunications.
Automatic Text Preprocessing for Intelligent Dialog Agents 809
The method used to decompose agglutinated verb-clitic forms has been inspired
by the work of [38,39]. The pronominal particles under examination have been
classified, for computational purposes, on the base of the number of character
by which they are composed:
• 2 characters: e.g. lo, la, mi, vi, ti; mangialo “eat it”, prenditi “take it”;
• 3 characters: e.g. llo, cci, mmi, gli; fallo “do it”, dacci “give us”, digli “tell
him”;
• 4 characters: e.g. melo, mela, visi; scrivimelo “write it to me”, dettamela
“dictate it to me”, sedervisi “to sit down there”;
• 5 characters: e.g. ccelo, ttele, mmene; daccelo “give it to us”, fattele “do them
by yourself”, fammene “do them to me”;
• 6 characters: e.g. glielo, glieli, gliene; daglielo “give it to him/her”, scrivigliene
“write him/her about it”, mandaglieli “send them to him/her”.
Starting from this classification, the system recognizes all the clitic particles
as the result of the Cartesian Product of basic elements. This way we can rapidly
generate all the possible combination (from which are excluded all the errors).
As an example, given two lists of basic elements,
a * b = [melo, mela, meli, mele, mene, mesi, meci, telo, tela, teli, tene, tesi,
teci, ...]
verbal stems listed in the dictionaries; only if the system finds a match, therefore
the candidate is suitable to become a true positive agglutinate form verb+clitic.
The module aims at reducing, as much as possible, computational mistakes,
such as homography and false positivity. Examples are the following:
• “dalla” in “dalla a me, la penna”, “give it to me, the pen”, where the token is
a true agglutinated form of “dare+la”; and “dalla finestra alla porta”, “from
the window to the door”, where the token is just a preposition “da” combined
with the definite article “la”);
• “mangiatene” in “mangiate-ne”, where the token agglutinates a participle
“mangiate” and the particle “ne”; and “mangia-te-ne” in which the token
agglutinates the imperative “mangia” and the particles “te” and “ne”.
With the first example, a syntactic analysis performed on the sentence could
disambiguate the agglutinated forms, but in the second case it will be impossible
to detect when the word is an agglutinated form or not, also with a deep context
analysis performed.
The effectiveness of the method has been tested on 10,000 sentences (500,623
tokens) which have been randomly extracted from the corpus Paisà [21]. The
module for the clitic decomposition recognized and analyzed 5,544 verbs and
5,496 verbal forms agglutinated with clitics. Table 2 shows the evaluation of the
module under examination, which has been compared with the performances
of two of the most used Italian parsers. The module emerges in term of Recall,
while it presents a small deficiency with the Precision measure. This is due to the
fact that the module works on single tokens, before the postagging is performed,
without considering the context of tokens. This choice affects the analysis of false
positives like “influenzale” (that is recognized by our system as an agglutinated
form verb + clitic particle “influenza-le”, “influence them”, when is actually an
adjective “flu-like”, correctly postagged by the other parsers.
However, the Recall value reached by our tool makes it obtain the best F1
score (0.996) with respect to LinguA (0,982) e Tint (0,915).
Table 2. Clitics decomposition module results compared with LinguA and TINT
results
3.3 Lemmatization
The Lemmatization module is based on the DELAF, the Italian Electronic Dic-
tionary of Flexed Forms. It includes over 1 million of inflected Italian forms and
it has been divided into 6 sub-dictionaries in order to improve performance of
the algorithm.
In our methodology we directly match tokens with the tags that have been
inherited during the postagging process. Therefore, the system input is a list of
couples of token-postagging descriptions. Each token is matched into a dictionary
which is specific for the part of speech of interest in the list.
The information about the token that the lemmatizer will insert into the
database are of these kinds:
Once that the token has been identified in the dictionary, comes the step of
the homonymy checking.
Dictionaries has been converted into a Json File in which each term is a
list. The first element of the list is the token, in order to easily extract all the
information once the POS Tagger analyze the word. The second element is a
Json Object in which all the morphosyntactical and semantic information about
the term is explicited. There are cases in which homonymy causes errors in the
lemmatization. However, it often happens that one of the results is obsolete (e.g.
sono that corresponds both to essere “to be” that sonare “to play”). In order to
quickly solve these problems we arranged a disambiguation list where we confer
a greater weight to the most common interpretations.
The Lemmatizator has been tested on a corpus of 250.000 tokens extracted
by a different portion of the “Paisá Corpus”. Due to the heterogeneous character
of the “Paisá Corpus” it is suitable for an efficient test phase.
To evaluate the module, we decided to compare its performances with Tree-
Tagger one of the most known tools for annotating text with part-of-speech and
lemma information for several languages.
We compared the tools evaluating the results obtained from the analysis on
the same data portion. An overview of the comparison is presented in Table 3.
POS Tag and Lemmatization module reaches very good results in terms of pre-
cision for lemmatization tasks. Conversely for what concerns the processing time,
the module still requires many adjustments. In fact, Tree-Tagger performs both
tasks concurrently and the processing time must be considered the same for POS
812 A. Maisto et al.
Tagging and Lemmatization. Otherwise, our module perform the two tasks subse-
quently, therefore the value must be added to the POS Tagging processing time.
At the end, we analyze the results in a qualitative way. The vast majority of
errors derives from an incorrect part-of-speech tagging in ambiguity cases (i.e.
condotta, V, lemmatized as condurre, “to lead”, instead of condotta, N that
could be lemmatized as condotta, N, “conduct”).
4 Conclusion
This research focused on the preprocessing step of a software which has been
developed for the linguistic interface of the BIG4MASS project. As this project
has been built around the Question Answering research topic, the linguistic
analysis phase is mandatory for the correct understanding of the questions and
the drafting of the answers. The preprocessing, first step in each text analysis,
could be crucial for the quality of the final results.
In our work we developed three main modules, which, starting from the style
correction, due to the characteristic of the domain corpora (in large part User
Generated Content), move thought the decomposition of agglutinated verb-clitic
forms, the POS Tagging and Lemmatization phase.
References
1. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for e-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). https://www.scopus.com/inward/record.uri?eid=2-s2.0-84949036166&doi=
10.1016\%2fj.jvlc.2015.10.012&partnerID=40&md5=77d8fbfd4293af91d382e77055
3e159c.
2. Amato, F., Mazzeo, A., Penta, A., Picariello, A.: Knowledge representation and
management for e-government documents. In: IFIP International Federation for
Information Processing, vol. 280, pp. 31–40 (2008). https://www.scopus.com/
inward/record.uri?eid=2-s2.0-47249122655&doi=10.1007%2f978-0-387-09712-1
4&partnerID=40&md5=d8e43aeb328d1fdcb45aa323689f7010.
3. Amato, F., Moscato, F.: A model driven approach to data privacy verifica-
tion in e-health systems. Trans. Data Privacy 8(3), 273–296 (2015). https://
www.scopus.com/inward/record.uri?eid=2-s2.0-84949775424&partnerID=40&
md5=2ca02581d8482fbc2f8489a423a80290.
4. Attardi, G., Fuschetto, A., Tamberi, F., Simi, M., Vecchi, E.M.: Experiments in
tagger combination: arbitrating, guessing, correcting, suggesting. In: Proceedings
of Workshop Evalita, p. 10 (2009)
Automatic Text Preprocessing for Intelligent Dialog Agents 813
5. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning
to align and translate. arXiv preprint arXiv:1409.0473 (2014)
6. Choi, J.D.: Dynamic feature induction: the last gist to the state-of-the-art. In:
Proceedings of NAACL-HLT, pp. 271–281 (2016)
7. Chrupala, G., Dinu, G., Van Genabith, J.: Learning morphology with Morfette
(2008)
8. Collins, M.: Discriminative training methods for hidden Markov models: theory
and experiments with perceptron algorithms. In: Proceedings of the ACL-02 Con-
ference on Empirical Methods in Natural Language Processing, vol. 10, pp. 1–8.
Association for Computational Linguistics (2002)
9. Constant, M., Tellier, I., Duchier, D., Dupont, Y., Sigogne, A., Billot, S.: Intégrer
des connaissances linguistiques dans un crf: application à l’apprentissage d’un
segmenteur-étiqueteur du français. In: TALN, vol. 1, p. 321 (2011)
10. Dahlmeier, D., Ng, H.T.: Better evaluation for grammatical error correction. In:
Proceedings of the 2012 Conference of the North American Chapter of the Associ-
ation for Computational Linguistics: Human Language Technologies, pp. 568–572.
Association for Computational Linguistics (2012)
11. Dell’Orletta, F.: Ensemble system for part-of-speech tagging. In: Proceedings of
EVALITA, vol. 9, pp. 1–8 (2009)
12. Delmonte, R.: Generating and parsing clitics with getarun. In: CLIN (1999)
13. Favretti, R.R., Tamburini, F., De Santis, C.: CORIS/CODIS: a corpus of written
Italian based on a defined and a dynamic model. A Rainbow of Corpora: Corpus
Linguistics and the Languages of the World. Munich: Lincom-Europa (2002)
14. Felice, M., Yuan, Z., Andersen, Ø.E., Yannakoudakis, H., Kochmar, E.: Grammat-
ical error correction using hybrid systems and type filtering. In: Proceedings of
the Eighteenth Conference on Computational Natural Language Learning: Shared
Task, pp. 15–24 (2014)
15. Giménez, J., Marquez, L.: Svmtool: a general POS tagger generator based on
support vector machines. In: Proceedings of the 4th International Conference on
Language Resources and Evaluation. Citeseer (2004)
16. Han, N.R., Chodorow, M., Leacock, C.: Detecting errors in English article usage
with a maximum entropy classifier trained on a large, diverse corpus. In: LREC
(2004)
17. Huang, Z., Xu, W., Yu, K.: Bidirectional LSTM-CRF models for sequence tagging.
arXiv preprint arXiv:1508.01991 (2015)
18. Ingason, A.K., Helgadóttir, S., Loftsson, H., Rögnvaldsson, E.: A mixed method
lemmatization algorithm using a hierarchy of linguistic identities (HOLI). In:
Advances in Natural Language Processing, pp. 205–216. Springer, Heidelberg
(2008)
19. Junczys-Dowmunt, M., Grundkiewicz, R.: The AMU system in the CoNLL-2014
shared task: grammatical error correction by data-intensive and feature-rich statis-
tical machine translation. In: Proceedings of the Eighteenth Conference on Com-
putational Natural Language Learning: Shared Task, pp. 25–33 (2014)
20. Levenshtein, V.I.: Binary codes capable of correcting deletions, insertions, and
reversals. Soviet Phys. doklady 10, 707–710 (1966)
21. Lyding, V., Stemle, E., Borghetti, C., Brunello, M., Castagnoli, S., Dell’Orletta,
F., Dittmann, H., Lenci, A., Pirrelli, V.: The PAISA corpus of Italian web texts.
In: Proceedings of the 9th Web as Corpus Workshop (WaC-9), pp. 36–43 (2014)
22. Manning, C.D.: Part-of-speech tagging from 97% to 100%: is it time for some
linguistics? In: International Conference on Intelligent Text Processing and Com-
putational Linguistics, pp. 171–189. Springer, Heidelberg (2011)
814 A. Maisto et al.
23. Mizumoto, T., Hayashibe, Y., Komachi, M., Nagata, M., Matsumoto, Y.: The
effect of learner corpus size in grammatical error correction of ESL writings. In:
Proceedings of COLING 2012: Posters, pp. 863–872 (2012)
24. Monachesi, P.: A grammar of Italian clitics (1996)
25. Monachesi, P.: A lexical approach to Italian cliticization (1999)
26. Palmero Aprosio, A., Moretti, G.: Italy goes to Stanford: a collection of CoreNLP
modules for Italian. ArXiv e-prints (2016)
27. Pianta, E., Zanoli, R.: TagPro: a system for Italian PoS tagging based on SVM.
Intelligenza Artificiale 4(2), 8–9 (2007)
28. Porter, M.F.: An algorithm for suffix stripping. Program 14(3), 130–137 (1980)
29. Rozovskaya, A., Roth, D.: Algorithm selection and model adaptation for ESL cor-
rection tasks. In: Proceedings of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies, vol. 1, pp. 924–933.
Association for Computational Linguistics (2011)
30. Russi, C.: Italian clitics: an empirical study, vol. 193. Walter de Gruyter (2008)
31. Schmid, H.: Treetagger—a language independent part-of-speech tagger. Institut
für Maschinelle Sprachverarbeitung, Universität Stuttgart 43, 28 (1995)
32. Shen, L., Satta, G., Joshi, A.: Guided learning for bidirectional sequence classifi-
cation. In: ACL, vol. 7, pp. 760–767. Citeseer (2007)
33. Smedt, T.D., Daelemans, W.: Pattern for python. J. Mach. Learn. Res. 13(Jun),
2063–2067 (2012)
34. Srinivasan, A., Compton, P., Malor, R., Edwards, G., Lazarus, L.: Knowledge
acquisition in context for a complex domain. Pre-print of Proceedings of the Fifth
EKAW91 (1991)
35. Sun, C., Jin, X., Lin, L., Zhao, Y., Wang, X.: Convolutional neural networks for
correcting English article errors. In: Natural Language Processing and Chinese
Computing, pp. 102–110. Springer, Heidelberg (2015)
36. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural
networks. In: Advances in Neural Information Processing Systems, pp. 3104–3112
(2014)
37. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech
tagging with a cyclic dependency network. In: Proceedings of the 2003 Conference
of the North American Chapter of the Association for Computational Linguistics on
Human Language Technology, vol. 1, pp. 173–180. Association for Computational
Linguistics (2003)
38. Vietri, S.: Dizionari elettronici e grammatiche a stati finiti: metodi di analisi for-
male della lingua italiana. Plectica (2008)
39. Vietri, S.: The Italian module for NooJ. In: Proceedings of the First Italian Con-
ference on Computational Linguistics, CLiC-it, pp. 389–393 (2014)
Exploiting Figures of Speech in Cultural
Heritage Reasoning
1 Introduction
Practically any asset of cultural heritage is drenched with metaphors and alle-
gories. For example, sovereigns have been represented or sculpted as mythologi-
cal figures. Gods where chosen to represent monarchs of various ages. Allegories
changed according to what was desired to be highlighted in the allegoric form:
Apollo-Sun was linked to sovereigns which wanted to emphasise the fact they
were the primary source of light for their kingdom and their subjects; Mars
exalted warrior and strategist properties of other monarchs (such as Napoleon).
Allegories in the iconography of sovereigns often were associated with their
lifestyle. Who was represented as the Sun God, maintained a more than lavish
life-style, who was painted as Mars was famous for his military victories. Still,
association to allegoric figures (always mythological ones) that were considered
negative.
Discovering hidden contents, authors provocations or clarifying unclear ele-
ments is easier when comparing elements in the asset domain with the metaphors
or allegories to which the elements are related. This is equivalent to reasoning in
abstract domain, and it is an old problem, but still investigated in literature [1,2].
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 815–821, 2019.
https://doi.org/10.1007/978-3-030-15035-8_79
816 F. Amato et al.
One of the first attempt of coupling metaphors and analysis of assets is in [3].
Anyway the work outlined only the need of extending automated reasoning to
metaphors. The idea of using formal semantics (by ontologies) in metaphorical
reasoning is really novel [2,4,5]. Authors of [5] reported a model where metaphors
are associated to attributes in ontologies defined by OWL. The works in [2,4]
contains one of the first attempt of formalizing metaphors by using ontology-
based approaches.
However, at the best of our knowledge, previous works do not address the
problem of linking domains represented in cultural heritages with the ones
related to metaphors or allegories. The approach we present here is the first
attempt to overcome this problem.
The same elements will be used while applying our approach, but it will
use ontology conceptualization of different knowledge domain in order to merge
ontologies by means metaphors. Metaphors identification will involve an con-
ceptualization of the cognitive approaches that lead to cope together symbols
from different knowledge domain. The methodology, which uses metaphors, will
involve interaction with formal semantics and semiotics domains.
It is clear that, in addition to the traditional tools for the study of social
science, Semantics (of documents) and Semiotics (a systemic relationship of signs
within communication processes) have to be used.
For historical and social sciences, the approach to the text must take into
account the relations between the languages and the speakers: words change their
meanings through the centuries and sometimes semantics related to phrases may
be lost. Researchers must take into account possible replacements of words and
meanings in order to re-build original semantics from historical documents.
The concept of meaning is important in linguistic theories. Each meaning is
defined only in relation to other meanings, within a given domain.
Exploiting Figures of Speech in Cultural Heritage Reasoning 817
The idea of matching both semantic and symbolic terminology for the descrip-
tion of the content of documents or archival artefacts opens new doors for what
regards their analysis. The allegoric figure, and generally that of the metaphor,
joins a meaning to a description when the description adopts terms and symbols
of a different domain to that of the intended meaning. If both the descriptive
and the metaphoric domain are described through ontologies, the first thing
that comes to mind is to associate, to concepts in the descriptive domain, those
of the metaphorical domain and, therefore, reason by analogy on the model in
the metaphorical domain, by adding some of the reasoning done to that of the
descriptive domain. It would be very useful to reason on documents by involv-
ing both domains. This way, connections could surface between concepts in the
descriptive domain that are difficult to understand if the reasoning was limited
to that domain, while are instead easy to highlight in the metaphorical domain.
Thanks to the use of formal (ontological) models [9–11] for the annotation of
documents, it is possible to introduce totally automatic reasoning processes,
namely create engines of artificial reasoning allowing to connect two domains of
interest and draw conclusions from their conjoined analysis.
One of the main reasons why an ontology is defined in OWL is to allow to
describe the content of a document or of an object in general, referring to some
sort of evolved vocabulary (that being the ontology), on which the most part of
users and scholars of such objects uniquely converge in the definition and use.
An Annotation is the link of ontologies elements to assets in cultural heritages
(whatever is the format used to store them: images, videos, digital documents
etc.) [12–15].
818 F. Amato et al.
For what regards annotations, there are different approaches. They can be
written directly into the document or through an external support (such as
another digital document); it may or may not allow the annotation through
individuals etc.
This is not the place to discuss all of the options, methodologies and tech-
niques used in the process of semantic annotation. A wide description can be
recovered from among the scientific literature.
What matters is that on the one hand domain experts are busy defining
ontologies to use within their scope of application; on the other, expert archivists,
historians, other humanities scholars, scientists etc. are working on connect-
ing the content of analysed documents to the concepts that are present within
domain ontologies. It is also clear that one document can be annotated for sev-
eral and different domain ontologies if the document is about concepts present
in more than one domain.
It is through annotation processes, and the analysis of the very annotations,
that it is possible to implement actions of semantic research within documents.
The procedures of semantic research are however difficult to apply to Web
documents, considering how many and how heterogeneous they are. They do
work well when related to narrow domains (or to a single semantic domain)
allowing for particular reasoning [16].
4 Metaphoric Abstraction
Metaphors are the basics elements of the methods we are goings to explain. The
definition of Metaphor, in the Oxford Dictionary, is the following: A figure of
speech in which a word or phrase is applied to an object or action to which it
is not literally applicable, or:A thing regarded as representative or symbolic of
something else.
The last definition is important because it contains the concept that a
Metaphor is something having a given semantics in a domain A, that is used to
address something else in an another domain B. The use, in the same document,
of usual concepts and metaphors, doubtless increases the complexity of the docu-
ment, but let us consider the case of a well-known, well-defined and clear domain
M, and a set of (historical) documents, from different periods or ages, belonging
to a knowledge domain D and that we suppose ‘metaphor-pruned’ (i.e. without
content containing metaphoric meanings). If we are able to describe semantics
of source documents in the domain D, in terms of the concepts of the domain M,
regardless of the meaning (i.e. semantics) of terms in the documents, we should
have the semantics of elements in the documents represented in terms of a clear
set of symbols (in his case words) with defined semantics [17–19]. Reasoning on
Domain M is simpler than reasoning on D and new concepts can be discovered
from source documents. More in detail, Abstraction, Pattern Recognition and
Generalisation are applied in our CT methodology. First of all, we have a set of
source documents (S) to analyse. Let us suppose that the documents belong to
different periods and that a domain ontology D exists able to describe the main
Exploiting Figures of Speech in Cultural Heritage Reasoning 819
concepts in the documents. Let us assume that terms and concepts related to
them have changed during the years so D is no more able to describe all the
meanings of the documents since some of them have been lost due to language
changes (Fig. 1).
The first pass of the method is the retrieval of suitable Metaphors for doc-
uments contents. Let us characterise the Metaphors in ontological way and let
us assume that M is the ontology used for defining meanings of concepts in the
metaphoric domain. Both D and M have to be defined with a formal language.
It is extremely important that M is the more possible detailed. In our example
we use OWL in order to define ontologies. The next steps are the following: 1.
Documents are annotated by using concepts in D ontology. 2. The Abstraction
technique is applied: If a metaphor is known linking concepts (and instances)
from D to M, the link between concepts is formally reported. This like an oper-
ation of Ontology Merging, but D and M remain separated. At the end of this
step, it is possible to refer to known concept in D by using a metaphoric symbol.
3. The Pattern Recognition technique is then applied: If (graph) patterns exist
in M but there is no related pattern in D, even if similar relations (in terms
of graph structure) exist in the element of D metaphoric linked to elements in
M, probably the metaphoric connection has been lost during the years. 4. The
Generalisation technique is applied when the lost metaphoric link are re-applied
to source documents, recovering lost semantics [20–22] and information.
References
1. Narayanan, S.: Moving right along: a computational model of metaphoric reasoning
about events. In: AAAI/IAAI, pp. 121–127 (1999)
2. Diarra, D., Clouzot, M., Nicolle, C.: Causal reasoning and symbolic relationships
in medieval illuminations (2018)
3. Endert, A., North, C., Chang, R., Zhou, M.: Toward usable interactive analyt-
ics: coupling cognition and computation. In: Proceedings of the ACM SIGKDD
Workshop on Interactive Data Exploration and Analytics (IDEA 2014), pp. 52–56
(2014)
4. Gangemi, A., Alam, M., Presutti, V.: Amnestic forgery: an ontology of conceptual
metaphors. arXiv preprint arXiv:1805.12115 (2018)
5. Bulat, L., Clark, S., Shutova, E.: Modelling metaphor with attribute-based seman-
tics. In: Proceedings of the 15th Conference of the European Chapter of the Asso-
ciation for Computational Linguistics, Short Papers, vol. 2, pp. 523–528 (2017)
6. Posch, A., Scholz, R.W.: Transdisciplinary case studies for sustainability learning.
Int. J. Sustain. High. Educ. 7(3) (2006)
7. Eco, U.: Dall’albero al labirinto: studi storici sul segno e l’interpretazione. La Nave
di Teseo Editore spa, Milan (2017)
8. Eco, U.: Trattato di semiotica generale. La Nave di Teseo Editore spa, Milan (2016)
9. Amato, F., Moscato, F., Moscato, V., Colace, F.: Improving security in cloud by
formal modeling of IaaS resources. Future Gener. Comput. Syst. 87, 754–764 (2018)
10. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. In: International Conference on Data Compression,
Communication, Processing and Security, CCPS-2016 (2016)
11. Cilardo, A.: Exploring the potential of threshold logic for cryptography-related
operations. IEEE Trans. Comput. 60(4), 452–462 (2011)
12. Piccialli, F., Chianese, A.: The internet of things supporting context-aware com-
puting: a cultural heritage case study. Mob. Netw. Appl. 22(2), 332–343 (2017)
13. Chianese, A., Marulli, F., Piccialli, F., Valente, I.: A novel challenge into mul-
timedia cultural heritage: an integrated approach to support cultural informa-
tion enrichment. In: 2013 International Conference on Signal-Image Technology &
Internet-Based Systems (SITIS), pp. 217–224. IEEE (2013)
14. Piccialli, F., Jung, J.E.: Understanding customer experience diffusion on social
networking services by big data analytics. Mob. Netw. Appl. 22(4), 605–612 (2017)
15. Amato, F., Moscato, V., Picariello, A., Piccialli, F.: SOS: a multimedia recom-
mender system for online social networks. Future Gener. Comput. Syst. (2017)
16. Mangold, C.: A survey and classification of semantic search approaches. Int. J.
Metadata Semant. Ontol. 2(1), 23–34 (2007)
17. Xhafa, F., Barolli, L.: Semantics, intelligent processing and services for big data.
Future Gener. Comput. Syst. 37, 201–202 (2014)
18. Moore, P., Xhafa, F., Barolli, L.: Semantic valence modeling: emotion recognition
and affective states in context-aware systems. In: 2014 28th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA),
pp. 536–541. IEEE (2014)
19. Xhafa, F., Asimakopoulou, E., Bessis, N., Barolli, L., Takizawa, M.: An event-
based approach to supporting team coordination and decision making in disaster
management scenarios. In: 2011 Third International Conference on Intelligent Net-
working and Collaborative Systems (INCoS), pp. 741–745. IEEE (2011)
Exploiting Figures of Speech in Cultural Heritage Reasoning 821
20. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for e-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016)
21. Balzano, W., Vitale, F.: DiG-Park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: 2017 31st International Conference
on Advanced Information Networking and Applications Workshops (WAINA), pp.
698–703. IEEE (2017)
22. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
Textual Processing in Social Network
Analysis
1 Introduction
The number of active users of social networking systems is amazing. Such kind
of systems can process and publish information with different formats. The pub-
lished information results inherently heterogeneous. Social media networks have
emerged as powerful means of communication for people looking to share and
exchange information on a wide variety of real-world events. These events range
from popular, widely known ones (e.g., a concert by a popular music band) to
smaller scale, local events (e.g., a local social gathering, a protest, or an acci-
dent). Short messages posted on social media sites such as Twitter can typically
reflect these events as they happen. For this reason, the content of such social
media sites is particularly useful for real-time identification of real-world events.
Social networks have become a virtually unlimited source of knowledge, that
can be used for scientific as well as commercial purposes. In fact, the analysis
of on-line public information collected by social networks is becoming increas-
ingly popular, not only for the detection of particular trend, close to the classic
market research problem, but also to solve problems of different nature, such
as the identification of fraudulent behaviour, optimization of web sites, tracking
the geographical location of particular users or, more generally, to find mean-
ingful patterns in a certain set of data. In particular, the rich knowledge that
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 822–832, 2019.
https://doi.org/10.1007/978-3-030-15035-8_80
Textual Processing in Social Network Analysis 823
has accumulated in Twitter enables to catch the happening of real world events
in real-time. These event messages can provide a set of unique perspectives,
regardless of the event type [1–3], reflecting the points of view of users who are
interested or even participate in an event. In particular, for unplanned events
(e.g., the Iran election protests, earthquakes), Twitter users sometimes spread
news prior to the traditional news media [4]. Even for planned events, Twitter
users often post messages in anticipation of the event, which can lead to early
identification of interest in these events. Additionally, Twitter users often post
information on local, community-specific events where traditional news coverage
is low or nonexistent. Thus Twitter can be considered as a collector of real-time
information that could be used by public authorities as an additional informa-
tion source for obtaining warnings on event occurrence. In the last few years,
particular interest has given to the extraction and analysis of information from
social media by authorities and structures responsible for the protection of public
order and safety. Increasingly, through the information posted on social media,
public authorities are able to conduct a variety of activities, such as prevention
of terrorism and bio-terrorism, prevention of public order problems and safety
guarantee during demonstrations with large participation of people.
The heterogeneity of information and the huge scale of data makes the identi-
fication of events from Twitter a challenging problem. In fact, Twitter messages,
or tweets, has a variety of content types, including personal updates, not related
to any particular real-world event, information about event happenings, retweets
of messages which are of interest for a user and so on [5]. As an additional chal-
lenge, Twitter messages contain little textual information, having by design the
limit of 140 characters and often exhibit low quality [6]. Several research efforts
have focused on identifying events in Twitter [7,8]. For example, in recent years
there has been a lot of research efforts in analysing Tweets to enhance health
related alerts [9], following a natural calamity or a bio-terrorist attack, which
can urge a rapid response from health authority, as well as disease monitoring
for prevention. The previous work in this area includes validating the timeliness
of Twitter by correlating Tweets with the real-world statistics of disease activ-
ities, e.g., Influenza-like-Illness (ILI) rates [10–13], E-coli [14], cholera [15] or
officially notified cases of dengue [16]. In this work, we present a general frame-
work for event detection, starting from the heterogeneous processing of data
coming from social networks systems [17–20]. The proposed system process het-
erogeneous information [21–23] in order to detect anomalies in Twitter stream.
Event related anomalies are patterns in data that do not fit the pattern of the
expected normal behaviour [24]. Those anomalies might be induced in the data
for malicious activity, as for example cyber-intrusion or terrorist activity. The
proposed framework aims at analysing tweets and automatically extract relevant
event related information in order to raise a signal as soon as a malicious event
activity is detected. Ad-hoc hardware devices and high-performance communi-
cation scheduling can be adopted to manage big amount of data [25–28].
The reminder of the paper is structured as follows: in Sect. 2 we describe
two examples that motivate our work; in Sect. 3 we present an overview of the
824 F. Amato et al.
framework along with the goals and requirements that are of importance for
event detection from Twitter. Moreover in this Section we describe the main
processing stages of the system. In the end, in Sect. 4 we present some conclusions
and future work.
2 Motivating Example
1
http://www.gallup.com/.
Textual Processing in Social Network Analysis 825
3 Framework Description
In this work we present a general framework to be adopted for event detection
from Twitter. In particular, we aim at detecting anomalies in Twitter stream
related to malicious activity, as for example cyber-intrusion or terrorist activ-
ity. In this section we present an overview of the main processing stages of the
Fig. 1. Overview of the twitter stream processing framework for event detection and
alerting.
Textual Processing in Social Network Analysis 827
framework along with the goals and requirements that are of importance for
event detection from Twitter. Event detection from Twitter messages must effi-
ciently and accurately filter relevant information about events of specific interest,
which is hidden within a large amount of insignificant information. The pro-
posed framework aims at analyzing tweets and automatically extract relevant
event related information in order to raise a signal as soon as a malicious event
activity is detected. It is worth noting that in this work we consider an event
occurrence in Twitter stream whenever an anomaly bursts in the analyzed data
stream. The Twitter stream processing pipeline consists of three main stages,
as illustrated in Fig. 1. The system aims at detecting user defined phenomenon
form the twitter stream in order to give alerts in case of event occurrence. In the
following we describe the main stages of the proposed system.
In our work we are interested in detecting specified event which relies on spe-
cific information and features that are known about the event such as type and
description, which are provided by the a domain expert of the event context.
These features are exploited by adapting traditional information extraction tech-
niques to the Twitter messages characteristics. In the first stage, the tweets are
collected exploiting the REST API provided from Twitter that allow program-
matic access to read data from the stream. Those API are costumed in order to
crawl data exploiting user defined rules (keywords, hashtag, user profile, etc.).
The tweets are annotated, in the filter module, with locations and temporal
expressions using a series of language processing tools for tokenization, part-
of-speech tagging, temporal expression extraction, and tools for named entity
recognition.
In order to access to Twitter data programmatically, it’s necessary to regis-
ter a trusted application, associated with an user account, that interacts with
Twitter APIs. Registering a trusted application, Twitter provides the required
credentials (consumer keys) that can be used to authenticate the REST calls via
the OAuth protocol. To create and manage REST calls we integrate in the Filter
Module a Python wrapper for the Twitter API, called Tweepy, that includes a
set of class enabling the interaction between the system and Twitter Stream.
For example, the API class provides access to the entire twitter RESTful API
methods. Each method can accept various parameters and return responses.
To gather all the upcoming tweets about a particular event, we call the
Streaming API extending Tweepy’s StreamListener() class in order to customise
the way we process the incoming data. In the Listing 1 we simply gather all the
new tweets with the #anthrax hashtag:
Listing 1. Custom Listener to gather all the new tweets with the #python hashtag
and save them in a file
class TwitterListener ( StreamListener ) :
def o n d a t a ( s e l f , data ) :
try :
with open ( ’ anthrax . j s o n ’ , ’ a ’ ) a s f :
f . w r i t e ( data )
return True
except BaseException as e :
print ( ” E r r o r o n d a t a : %s ” % s t r ( e ) )
return True
Textual Processing in Social Network Analysis 829
def o n e r r o r ( s e l f , s t a t u s ) :
print ( s t a t u s )
return True
t w i t t e r s t r e a m = Stream ( auth , T w i t t e r L i s t e n e r ( ) )
t w i t t e r s t r e a m . f i l t e r ( t r a c k =[ ’#anthrax ’ ] )
All gathered tweets will be furthermore filtered and then passed to the Clas-
sification module for deeper textual processing.
The collected and annotated tweets are then classified, in the second stage, with
respect to their relevance. As stated before in the paper, event detection from
Twitter messages is a challenging task since the relevant information about a
specific event is hidden within a large amount of insignificant messages. Thus a
classifier must efficiently and accurately filter relevant information. The classifier
is trained exploiting a domain dependent thesaurus that include a list of relevant
terms for event detection. It is responsible for filtering out irrelevant messages.
Once the set of irrelevant tweets are discarded, those remaining must be con-
veniently aggregated with respect to an event of interest. The third stage of
the proposed system is aimed at detecting anomalies in the event related set
of tweets. Anomalies are patterns in data that do not fit the pattern of the
expected normal behavior. Anomalies might be induced in the data for a variety
of reasons, such as malicious activity, as for example cyber-intrusion or terror-
ist activity, but all of the reasons have a common characteristic that they are
interesting to the analyst. In literature there are several techniques exploited for
the anomaly detection. Those range between several disciplines and approaches
such as statistics, machine learning, data mining, information theory, spectral
theory.
In our proposed system, the set of event related tweets are processed by the
alert module that is responsible to raise an alert in case of anomaly within the
collected tweets. The alert module implements a set of burst detection algorithms
in order to generate alerts whenever an event occurs that means an anomaly in
the analyzed data stream is detected. The detected events are then displayed to
the end user exploiting the visualization module. An important aspect is that
the analyzed data can be tuned by the end user for novelty detection which aims
at detecting previously unobserved (emergent, novel) patterns in the data, such
as a new topic. The novel patterns are novel patterns are typically incorporated
into the normal model after being detected. Moreover with the tuning phase
the parameters for the tweets classification module can be modified in case the
messages that raised an alert are considered not relevant for the event occurrence.
830 F. Amato et al.
References
1. Diakopoulos, N., Naaman, M., Kivran-Swaine, F.: Diamonds in the rough: social
media visual analytics for journalistic inquiry. In: 2010 IEEE Symposium on Visual
Analytics Science and Technology (VAST), pp. 115–122, October 2010
2. Yardi, S., Boyd, D.: Tweeting from the town square: measuring geographic local
networks. In: International Conference on Weblogs and Social Media. American
Association for Artificial Intelligence, May 2010
3. Amato, F., Cozzolino, G., Mazzeo, A., Moscato, F.: Detect and correlate informa-
tion system events through verbose logging messages analysis. Computing (2018).
Cited By 0; Article in Press
4. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes Twitter users: real-time
event detection by social sensors. In: Proceedings of the 19th International Con-
ference on World Wide Web, WWW 2010, pp. 851–860. ACM, New York (2010)
5. Naaman, M., Boase, J., Lai, C.-H.: Is it really about me?: Message content in social
awareness streams. In: Proceedings of the 2010 ACM Conference on Computer
Supported Cooperative Work, CSCW 2010, pp. 189–192. ACM, New York (2010)
6. Becker, H., Naaman, M., Gravano, L.: Beyond trending topics: real-world event
identification on Twitter (2011)
7. Becker, H., Naaman, M., Gravano, L.: Learning similarity metrics for event iden-
tification in social media. In: Proceedings of the Third ACM International Con-
ference on Web Search and Data Mining, WSDM 2010, pp. 291–300. ACM, New
York (2010)
8. Atefeh, F., Khreich, W.: A survey of techniques for event detection in Twitter.
Comput. Intell. 31(1), 132–164 (2015)
9. Essmaeel, K., Gallo, L., Damiani, E., De Pietro, G., Dipanda, A.: Comparative
evaluation of methods for filtering kinect depth data. Multimedia Tools Appl.
74(17), 7331–7354 (2015)
Textual Processing in Social Network Analysis 831
10. Aramaki, E., Maskawa, S., Morita, M.: Twitter catches the flu: detecting influenza
epidemics using Twitter. In: Proceedings of Conference on Empirical Methods in
Natural Language Processing (2011)
11. Lamb, A., Paul, M.J., Dredze, M.: Separating fact from fear: tracking flu infections
on Twitter. In: HLT-NAACL, pp. 789–795 (2013)
12. Chon, J., Raymond, R., Wang, H., Wang, F.: Modeling flu trends with real-time
geo-tagged Twitter data streams. In: Wireless Algorithms, Systems, and Applica-
tions, pp. 60–69. Springer, Heidelberg (2015)
13. Amato, F., Cozzolino, G., Mazzeo, A., Romano, S.: Intelligent medical record man-
agement: a diagnosis support system. Int. J. High Performance Comput. Netw.
12(4), 391–399 (2018)
14. Diaz-Aviles, E., Stewart, A.: Tracking Twitter for epidemic intelligence: case study:
Ehec/hus outbreak in Germany. In: Proceedings of the 4th Annual ACM Web
Science Conference, WebSci 2012, pp. 82–85. ACM, New York (2012)
15. Chunara, R., Andrews, J.R., Brownstein, J.S.: Social and news media enable esti-
mation of epidemiological patterns early in the 2010 Haitian cholera outbreak. Am.
J. Trop. Med. Hyg. 86(1), 39–45 (2012)
16. Gomide, J., Veloso, A., Meira Jr, W., Almeida, V., Benevenuto, F., Ferraz, F.,
Teixeira, M.: Dengue surveillance based on a computational model of spatio-
temporal locality of Twitter. In: Proceedings of ACM WebSci 2011 (2011)
17. Piccialli, F., Chianese, A.: The internet of things supporting context-aware com-
puting: a cultural heritage case study. Mob. Netw. Appl. 22(2), 332–343 (2017)
18. Chianese, A., Marulli, F., Piccialli, F., Valente, I.: A novel challenge into mul-
timedia cultural heritage: an integrated approach to support cultural informa-
tion enrichment. In: 2013 International Conference on Signal-Image Technology &
Internet-Based Systems (SITIS), pp. 217–224. IEEE (2013)
19. Piccialli, F., Jung, J.E.: Understanding customer experience diffusion on social
networking services by big data analytics. Mob. Netw. Appl. 22(4), 605–612 (2017)
20. Amato, F., Moscato, V., Picariello, A., Piccialli, F.: SOS: a multimedia recom-
mender system for online social networks. Future Gen. Comput. Syst. (2017)
21. Xhafa, F., Asimakopoulou, E., Bessis, N., Barolli, L., Takizawa, M.: An event-
based approach to supporting team coordination and decision making in disaster
management scenarios. In: 2011 Third International Conference on Intelligent Net-
working and Collaborative Systems (INCoS), pp. 741–745. IEEE (2011)
22. Xhafa, F., Barolli, L.: Semantics, intelligent processing and services for big data
(2014)
23. Moore, P., Xhafa, F., Barolli, L.: Semantic valence modeling: emotion recognition
and affective states in context-aware systems. In: 2014 28th International Confer-
ence on Advanced Information Networking and Applications Workshops (WAINA),
pp. 536–541. IEEE (2014)
24. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection: a survey. ACM Com-
put. Surv. 41(3), 15:1–15:58 (2009)
25. Cilardo, A.: Exploring the potential of threshold logic for cryptography-related
operations. IEEE Trans. Comput. 60(4), 452–462 (2011)
26. Cilardo, A., Fusella, E., Gallo, L., Mazzeo, A.: Joint communication scheduling and
interconnect synthesis for FPGA-based many-core systems. In: Design, Automa-
tion and Test in Europe Conference and Exhibition (DATE), pp. 1–4. IEEE (2014)
27. Cilardo, A., Fusella, E., Gallo, L., Mazzeo, A.: Exploiting concurrency for the
automated synthesis of MPSoC interconnects. ACM Trans. Embedded Comput.
Syst. (TECS) 14(3), 57 (2015)
832 F. Amato et al.
28. Cilardo, A., Durante, P., Lofiego, C., Mazzeo, A.: Early prediction of hardware
complexity in HLL-to-HDL translation. In: 2010 International Conference on Field
Programmable Logic and Applications (FPL), pp. 483–488. IEEE (2010)
29. Hobbs, J.R., Riloff, E.: Information extraction. In: Indurkhya, N., Damerau, F.J.
(eds.) Handbook of Natural Language Processing, 2nd edn. CRC Press, Taylor and
Francis Group, Boca Raton (2010)
30. Butler, C.S.: Statistics in Linguistics. Blackwell, Oxford (1985)
31. Biber, D., Reppen, R., Conrad, S.: Corpus Linguistics: Investigating Language
Structure and Use. Cambridge University Press, Cambridge (1998)
32. Kennedy, G.D.: An Introduction to Corpus Linguistics. Longman, London (1998)
33. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. In: The International Conference on Data Compression,
Communication, Processing and Security, CCPS 2016 (2016)
34. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
35. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
36. Riloff, E.: Automatically constructing a dictionary for information extraction tasks.
In: Proceedings of the Eleventh National Conference on Artificial Intelligence
(1993)
37. Grishman, R.: Information extraction: Capabilities and challenges (2012)
38. Pantel, P., Pennacchiotti, M.: Espresso: leveraging generic patterns for automati-
cally harvesting semantic relations. In: Proceedings of the 21st International Con-
ference on Computational Linguistics and the 44th Annual Meeting of the Associ-
ation for Computational Linguistics (2006)
ACOp: An Algorithm Based on Ant
Colony Optimization for Parking Slot
Detection
1 Introduction
VANET has become an active research area due to its huge potential to improve
vehicle and road safety [4], and traffic control [6,27]. In such a network, vehicles
act as nodes of a graph where connections represent communication between
nodes: if two vehicle are able to communicate (i.e. they are in the respective
communication ranges), there is a connection between the corresponding nodes
[13]. Each vehicle can exchange network information with its direct neighbors,
while through multi-hop communication further vehicles can be reached [11,12].
Moreover, information exchange may be performed not only between vehicles,
but also between vehicles and some road infrastructure items, such as a traffic
light: indeed, typically two kind of communications are possible in a VANET:
Vehicle-to-Vehicle and Vehicle-to-Infrastructure [16,19].
With the purpose of improving traffic control in a VANET, in this work we
focus on parking detection. Indeed, the presence of wandering vehicles looking
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 833–840, 2019.
https://doi.org/10.1007/978-3-030-15035-8_81
834 W. Balzano and S. Stranieri
for a free parking slot can significantly impact the traffic congestion in a VANET
and, beyond that, it constitutes a waste of time for drivers.
Automatic parking systems are made of different components, such as (i )
environment recognition, (ii ) path generation, and (iii ) path tracking [14,22].
In this work, we focus first on environment recognition phase, by providing a
structural representation of the considered area through a graph representing
the environment by regions of parking slot, and then on generation phase, by
applying ant colony optimization (ACO) to plan the optimal path leading to a
free parking slot. Our aim is to exploit a modeling approach, as in [5], whose
semantic [3] can be exploited to facilitate parking detection.
ACO is a paradigm for optimization algorithms that combines a priori knowl-
edge about promising solutions with a posteriori information of good solutions
[25]. This paradigm allows to build a complete, global, optimal solution through
successive approximations of good, local solutions.
The approach is bio-inspired, since it follows the behavior of ants in real
world: when ants look for food, they travel along the path having the highest
pheromone, which is the hormone secreted by other ants that have previously
taken that path [24].
Similarly, algorithms following this strategy are based on a loop performing
at each step opportune updates of global values in order to make a trail more
attractive than another one, according previous iterations results.
The ant colony optimization approach has been used already in literature to
address VANET challenges, such as clustering [9], and routing issues, but it has
never been applied for parking optimization purposes. The main contribution
of this work is a novel approach to the parking slot detection problem which
experiences a totally new combination with ant colony optimization.
Outline
The rest of the paper is organized as follows: in Sect. 2, an overview of existing
work is presented; in Sect. 3, we provide our ACOp approach to find the optimal
path with respect to the purpose of finding available parking slot; finally, in
Sect. 4, the conclusions are shown.
2 Related Work
Free parking slot detection is an interesting research field in VANETs context
[10,17], since it determines a big waste of time for drivers, and consequently
a traffic congestion increasing. For this reason, many proposals of automatic
parking systems are made by researchers trying to address the main VANET
issues, such as transmission delay, routing failure, and traffic congestion.
A lot of effort has been done by researcher to optimize network communica-
tion [7,8]. In [23], they point out that methods for target position-designation,
which is one of the core technologies in automatic parking systems, can be
divided into four categories: (i ) user-interface based, (ii ) parking slot mark-
ings based, (iii ) free-space based, and (iv ) infrastructure based. In [26], they
ACOp: An Algorithm Based on Ant Colony Optimization 835
propose a technique for vacant parking slot detection based on three stages: (i )
parking slot marking detection, (ii ) occupancy classification, and (iii ) parking
slot marking tracking. The first step recognizes different parking slot markings,
the second step identifies if the slots detected in the previous step are free, and
the third one estimates the position of the selected parking slot. Similarly, in
[22], they also provide an algorithm based on parking-slot-markings recognition.
They build a priori knowledge through one-dimensional filter in Hough space.
In our previous work [15], we analyze the distance geometry problem (DGP) and
we adopt it to the parking issue, by dividing vehicles in three groups according
to the possible states they can assume (parked, finding, and leaving), and man-
aging the interaction between these states in a probabilistic way through Markov
chains.
Ant Colony Optimization (ACO) is a paradigm for solving optimization prob-
lem that combines a priori information about promising solutions with a pos-
teriori information about previously obtained good solutions [25]. In literature,
we find different proposals applying ACO approach to VANET: in [1], they high-
lights that a VANET is more stable when an optimal number of clusters is built,
because network resources are efficiently employed, and for this reason they use
ant colony optimization applied to clustering techniques for VANETs. Results
show that ACO improves clustering performances, providing a technique which
is effective in comparison with others. Similarly, in [18], ACO is used to improve
VANET potentialities. In particular, their goal is to avoid the delay of com-
munication that occurs due to disconnection: for this reason, they apply ant
colony optimization to introduce repair algorithm in case of failure. In [24], they
use ACO to discover new routes in a VANET in an efficient way, with positive
results from performances point of view. In [21], they provide MACO, a modified
version of ant colony optimization, in which repulsion effect is used instead of
attractiveness in order to address the traffic congestion. In particular, they apply
ACO only in case of congestion, allowing vehicles to follow the non congested
path.
Despite the numerous application of ACO to vehicular context, to the best
of our knowledge, this is the first work addressing the free parking slot detection
by using bio-inspired ant colony optimization approach.
In this section, we want to propose a novel way to apply a family of solutions, the
Ant Colony Optimization, focusing on vehicular context, in particular parking
slot detection. Since free slot search constitutes a big waste of time for drivers,
many proposals to achieve good solutions to the problem have been done over
the last years.
Our goal is to formulate an optimization problem that expresses the need
of finding free parking slot in an efficient way, and then apply the ant colony
paradigm to solve that optimization problem.
836 W. Balzano and S. Stranieri
In this section, we want to classify the context of our problem [2]. Given the
problem of parking, the environment in which the problem can be imagined is
a huge area having taken or free parking slot. Our first assumption is to divide
the environment in regions, such that each region has a capability in terms of
available free parking slot.
The environment can be represented as a directed graph G = (V, E), where
each vertex v ∈ V is a region with |V | = n, and each edge (i, j) ∈ E is a
connection between regions, such (i, j) ∈ E if and only if from the region i the
region j is reachable directly. Moreover, a weight function w : V × V −→ N
associates to each edge a weight given by the capability of the destination node.
At any iteration, the ant k performs a move from i to j with a certain prob-
ability pkij that depends on attractiveness and trail level values. The probability
is computed as follows:
⎧ α β
⎪
⎨ τij + ηij
if (ij) ∈ tabuk
pkij = α + ηβ )
(τij ij
(1)
⎪
⎩
(ij) ∈ tabuk
0 otherwise
where:
• α and β are parameters between 0 and 1 that allow to specify how determinant
are the trail level and the attractiveness, respectively;
• tabuk is the list of not feasible moves for the ant k starting from the state i.
Once any ant completes its move, the trail level at iteration t is updated
according to the trail level at the previous iteration, as follows:
Fig. 1. Phases of a typical algorithm solving a problem according to ant colony opti-
mization
The updating phase is the moment when trail level and pheromone are mod-
ified: in particular, the trail level can either increase or decrease according to if
ants deposit pheromone or not.
In terms of the graph we have built in Subsect. 3.1, this corresponds to find a
path, from the region S where the driver starts to the region E where the driver
has to go to, with maximum weights over the edges.
Formally, let wij be the weight of the edge i → j in the graph G, and let
P be the resulting path of maximum weights, the optimization problem can be
formulated as follows: ⎧
n n
⎪
⎪
⎪
⎪ max wij ∗ xij
⎪
⎪
⎪
⎪ i=1 j=1
⎪
⎪ n
⎪
⎪
⎨ xij ≤ 1
(3)
⎪
⎪
i=1
⎪
⎪ n
⎪
⎪
⎪
⎪ xij ≤ 1
⎪
⎪
⎪
⎪ j=1
⎩
P [1] = S
where xij can assume values in [0, 1] according to the resulting path P :
1 if i → j ∈ P
xij =
0 otherwise
Fig. 2. Example of a graph of five regions and corresponding weighted connections (a),
and the chosen maximizing free parking slot path assuming A and E as starting and
ending point respectively (b).
The first inequality constraint guarantees that each node of the resulting
path has at most one outgoing arc (except for E that has not outgoing arcs).
Similarly, the second inequality constraint ensures that any node has at most
one incoming arc (zero in case of S). Notice that in the optimization problem
shown in (3), it is required that S appears in the resulting path (as first region
precisely), but this does not happen for E: indeed, if the destination region has
not free slot, any edge leading to it has a null weight and hence it is not chosen
for the optimal path.
ACOp: An Algorithm Based on Ant Colony Optimization 839
4 Conclusions
Ant colony optimization is largely used to solve known problems, such as TSP
(Traveling Salesman Problem), SOP (Sequential Ordering Problem), GAP (Gen-
eralized Assignment Problem), and MKP (Multiple Knapsack Problem) [20]. In
this work, we turn the free parking slot detection problem into an optimization
problem, with the aim of finding a feasible solution that maximize the objective
function. We first build an opportune environment representation, then we for-
mulate the optimization problem to find a path from a starting position, leading
as close as possible to a destination, that maximizes the number of free parking
slot seen. As far as we know, this is the first work relating ACO optimization
with parking problem.
References
1. Aadil, F., Bajwa, K.B., Khan, S., Chaudary, N.M., Akram, A.: CACONET: ant
colony optimization (ACO) based clustering algorithm for vanet. PloS One 11(5),
e0154080 (2016)
2. Amato, F., Boselli, R., Cesarini, M., Mercorio, F., Mezzanzanica, M., Moscato, V.,
Persia, F., Picariello, A.: Challenge: processing web texts for classifying job offers,
pp. 460–463 (2015). Cited By 16
3. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for e-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). Cited By 18
4. Amato, F., Mazzocca, N., Moscato, F.: Model driven design and evaluation of
security level in orchestrated cloud services. J. Netw. Comput. Appl. 106, 78–89
(2018). Cited By 2
5. Amato, F., Moscato, F.: Model transformations of mapreduce design patterns for
automatic development and verification. J. Parallel Distrib. Comput. 110, 52–59
(2017). Cited By 3
6. Balzano, W., Murano, A., Vitale, F.: V2V-en-vehicle-2-vehicle elastic network. Pro-
cedia Comput. Sci. 98, 497–502 (2016)
7. Balzano, W., Murano, A., Vitale, F.: WiFACT–wireless fingerprinting automated
continuous training. In: Proceedings of WAINA. IEEE Computer Society (2016)
8. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
9. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: 3PGCIC. Springer (2016)
10. Balzano, W., Vitale, F.: Dig-park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: Proceedings of the WAINA. IEEE
Computer Society (2017)
11. Balzano, W., Barbieri, V., Riccardi, G.: Car2Car framework based on DDGP3
12. Balzano, W., Del Sorbo, M.R., Stranieri, S.: A logic framework for C2C network
management. In: 2016 30th International Conference on Advanced Information
Networking and Applications Workshops (WAINA), pp. 52–57. IEEE (2016)
13. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
840 W. Balzano and S. Stranieri
14. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. In: International Journal of Grid and Utility Computing
(2017)
15. Balzano, W., Stranieri, S.: LoDGP: a framework for support traffic information
systems based on logic paradigm. In: International Conference on P2P, Parallel,
Grid, Cloud and Internet Computing, pp. 700–708. Springer (2017)
16. Balzano, W., Vitale, F.: DGP application for support traffic information systems
in indoor and outdoor environments. In: International Conference on P2P, Parallel,
Grid, Cloud and Internet Computing, pp. 692–699. Springer (2017)
17. Balzano, W., Vitale, F.: PAM-SAD: ubiquitous car parking availability model
based on V2V and smartphone activity detection. In: International Conference
on Intelligent Interactive Multimedia Systems and Services, pp. 232–240. Springer
(2017)
18. Chauhan, R.K., Dahiya, A.: AODV extension using ant colony optimization for
scalable routing invanets. J. Emerg. Trends Comput. Inf. Sci. 3(2), 241–244 (2012)
19. Da Cunha, F.D., Boukerche, A., Villas, L., Viana, A.C., Loureiro, A.A.: Data
communication in VANETs: a survey, challenges and applications. Ph.D. thesis,
INRIA Saclay; INRIA (2014)
20. Dorigo, M., Birattari, M., Blum, C., Clerc, M., Stützle, T., Winfield, A.: Ant
Colony Optimization and Swarm Intelligence: Proceedings of the 6th Interna-
tional Conference, ANTS 2008, Brussels, Belgium, 22–24 September 2008, vol.
5217. Springer (2008)
21. Jindal, V., Dhankani, H., Garg, R., Bedi, P.: MACO: modified ACO for reducing
travel time in VANETs. In: Proceedings of the Third International Symposium on
Women in Computing and Informatics, pp. 97–102. ACM (2015)
22. Jung, H.G., Kim, D.S., Yoon, P.J., Kim, J.: Parking slot markings recognition for
automatic parking assist system. In: 2006 IEEE Intelligent Vehicles Symposium,
pp. 106–113. IEEE (2006)
23. Jung, H.G., Lee, Y.H., Kim, J.: Uniform user interface for semiautomatic parking
slot marking recognition. IEEE Trans. Veh. Technol. 59(2), 616–626 (2010)
24. Majumdar, S., Rajendra Prasad, P., Santosh Kumar, S., Sunil Kumar, K.N., et al.:
An efficient routing algorithm based on ant colony optimisation for VANETs. In:
IEEE International Conference on Recent Trends in Electronics, Information and
Communication Technology (RTEICT), pp. 436–440. IEEE (2016)
25. Maniezzo, V., Gambardella, L.M., De Luigi, F.: Ant colony optimization (2004).
Cited in New Optimization Techniques in Engineering. Studies in Fuzziness and
Soft Computing, vol. 141. Springer, Heidelberg, p. 62 (2010)
26. Suhr, J.K., Jung, H.G.: Sensor fusion-based vacant parking slot detection and
tracking. IEEE Trans. Intell. Transp. Syst. 15(1), 21–36 (2014)
27. Zeadally, S., Hunt, R., Chen, Y.-S., Irwin, A., Hassan, A.: Vehicular ad hoc net-
works (VANETs): status, results, and challenges. Telecommun. Syst. 50(4), 217–
241 (2012)
A Self-organization Technique in Wireless
Sensor Networks to Address Node
Crashes Problem and Guarantee Network
Connectivity
1 Introduction
Wireless Sensor Networks (WSN) have lately attracted the attention of
researchers from different research fields, where modeling aimed to security
improvement are hot topics [2–4], due to their potential and the numerous appli-
cation areas, such as traffic control [14,16], road safety [7,11], object classification
[1] and tracking [5,13], environment monitoring [8,9,23], but also surveillance
[19] and parking [12].
This kind of networks have two main limitations: (i ) energy-constrained sen-
sors, and (ii ) limited communication range. In order to optimize the use of the
restricted bandwidth, a smart message transmission system is realized: a sen-
sor in a WSN is allowed to communicate only with its one-hop neighbor, while
long-range transmission are obtained by multi-hop communication, letting all
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 841–850, 2019.
https://doi.org/10.1007/978-3-030-15035-8_82
842 W. Balzano and S. Stranieri
the sensors cooperate to execute distributed tasks. Shorabi, in [26], gives a hier-
archical interpretation of a wireless sensor network, allowing to process data
and information at different abstraction levels: each event occurring in the net-
work environment can be processed at (i ) node level, meaning from the sensor
detecting the event point of view, (ii ) local neighborhood level, meaning from
the one-hop neighbors of the sensor detecting the event point of view, and (iii )
global level, which involves the whole network.
Sensors in WSN typically deal with highly dynamic environment: changes
in the weather conditions, occurrence of exceptional events, error detection are
some of the possible causes that determine rapid and frequent changes in the
environment conditions. In these cases, a WSN is supposed to be autonomous
enough to self-manage, without any human intervention [24]. Essentially, a WSN
is required to be fault-tolerant.
The ability of a WSN of solving problems after exceptional events by itself is
known as self-organizing property, meaning the autonomous organization of con-
nectivity, and data addressing. Self-organization is a challenge in wireless sensor
networks, because of the mentioned limitations like restricted energy and band-
width. The power constraint, especially, poses the problem of fault-tolerance:
when a sensor shuts down, because of energy expiration, the connectivity of the
whole network must not be affected by the failure of a single node.
In order to address the constrained energy of sensors, clustering techniques
are often employed to reduce power waste. In a typical scenario, data retrieved by
each sensor in a WSN are transmitted to central processing units, that gather all
the information and communicate them outside the network. As they point out
in [15], the cost for transmitting information is often higher than the one needed
for the computation. This fact suggests using clustering approaches. Indeed,
intuitively dividing sensors in groups reduces the distances that a single node
has to cover, because each sensor has to communicate its own information to
a representative node of its group, known as clusterhead, which is closer, in
terms of distances, than the central processing unit. The clusterheads of each
group receive data from all the other related sensors, and are responsible for
the transmission of such data to the central units. This mechanism reduces the
energy waste significantly for the whole network.
As we show in the next section, researchers have proposed different clustering
algorithms to optimize transmission in wireless sensor networks, and to make
the most of the available resources. Some of them also focused on sensors power
issue, providing energy-based clustering algorithms, but very few addressed the
problem of WSN fault-tolerance. The main contribution of this paper is a fault-
tolerant clustering algorithm, aiming to preserve the network connectivity even
after node crashes, and to improve the self-organizing ability.
Outline
The rest of the work is organized as follows: in Sect. 2, the state of the art is
analyzed and the existing clustering techniques are shown; in Sect. 3, a clustering
algorithm which is resistant to faults and node crashes is proposed, with relative
detection and recovery phases; finally, in Sect. 4, the conclusions of this work.
A Self-organization Technique in Wireless Sensor Networks 843
2 Related Work
One of the hot topics in Wireless Sensor Networks field is the self-organization
problem: as the authors point out in [24], WSNs should be able to manage
themselves and to work without any human intervention. Flako Dressler in [17]
explains that self-organization is a paradigm to handle distributed systems: this
property is required when the system need to face a change in the environment
circumstances, or to adapt to new scenarios, features that allow to improve
scalability and adaptivity in any control system.
In [26], the authors provide a survey explaining the existing self-organization
techniques, with related progress and open issues. They highlight that the WSN
ability to self-organize has been emphasized to address the problem of managing
large-scale networks, and has been proven to bring significant benefits in coop-
erative systems performances. On the other hand, self-organization requires a
new control theory to realize such an autonomous system.
For this reason, many researchers propose clustering as control strategy to
achieve such a self-managing ability. In our previous work, we focus on clus-
terization too [6,10]. In [22], they propose a load-balanced clustering algorithm
based on distance and density distribution (DSBCA). The aim of this algorithm
is to guarantee the generation of clusters that have balanced energy and to avoid
the situation with overcrowding clusters. Essentially, they propose to compute
the clustering radius based on density and distance such that when two clusters
have same density, the much further one from the base station takes a larger
radius, and when two clusters have same distance from the base station, the one
having higher density takes a smaller radius.
In [25], they exploit the fact that in a wireless sensor network communication
is the most consuming operation in terms of energy and they provide a set of
algorithms for keeping the network connected: these algorithms are based on
protocols for self-organization.
Authors in [21] show three different approaches to achieve self-organizing
property, by exploiting clustering algorithms to create groups of fixed size B : (i )
in the Expanding Rapid Alorithm, an initiator node includes in its cluster the
sensors in a neighborhood that increases at each iteration, until the cluster-size
is reached (some nodes of the last level may be cut off in case B is exceeded);
(ii ) in the Rapid Algorithm, the initiator, instead, is assigned a budget (given by
the cluster-size) which is first decremented and then equally distributed among
its one-hop-neighbors, that will repeat the procedure until the budget is expired
or there is no possibility to let the cluster grow: the main drawback is that this
algorithm may produce groups smaller than the resource would allow; (iii ) for
this reason, they propose a third algorithm, the Persistent Algorithm, which is
an extension of the previous one since it allows the nodes receiving the budget
to give back to the father the unused part of the budget, so that it can be
distributed again: this guarantees that the size of the groups is much more close
to the maximum cluster-size.
Moreover, in [15], they propose a randomized, distributed clustering algorithm
to manage sensors in WSNs and then they extend it to generate a hierarchy
844 W. Balzano and S. Stranieri
cluster creation allows a smart information distribution and data collection, since
there is a single node, elected as cluster-head, which is responsible for gathering
the information about all the other sensors in the same cluster, and transmitting
it outside the network. Clearly, dividing a network of hundreds of sensors in
groups of tens reduces not only the network congestion but also the effort made
by each node to collect all the data sent in the network.
In order to guarantee network connectivity, we propose a clustering algo-
rithm, based on backup-clusters, that helps the network be resistant with respect
to errors. A cluster is a connected component, meaning that there exists a path
between any pair of nodes, regardless of the cluster-head choice: this fact guar-
antees that, even after the cluster construction, the role of clusterhead might
change without impact the cluster connectivity.
Anytime a node appears in the network, it waits a random time t during which
it can be asked to join an existing cluster. If this time expires without any
invitation, the node becomes a temporary cluster-head, and it invites its adjacent
nodes to join the cluster. Given the considerations made previously about cluster
connectivity, during the cluster composition the cluster-head role can be given
to another node if it has left more energy. This choice aims to facilitate the
recovery after a crash: clearly, if an error occurs in the cluster-head the recovery
is more difficult than it would be if any other node turns off, as we will explain
later.
Let’s suppose to have a fixed maximum cluster-size S : this size is used during
the cluster creation as criterion to stop the cluster expansion, but it will be
relaxed with a tolerance ε during the recovery.
During the clustering procedure, each node is assigned:
846 W. Balzano and S. Stranieri
• a cluster id, which identifies its cluster, and thus the corresponding cluster-
head;
• a backup-clusters list, given by the cluster ids of adjacent nodes not belonging
to the same cluster of the current one. It would be appropriate keeping this list
ordered according to the backup-clusters emptiness: this requirement allows
to avoid overcrowding during the recovery, that would reduce significantly
the benefits of clusterization. In this way, the first item of the list contains
the most available backup-cluster id.
At any time a node is aware of its current cluster and, if a fault determining
its isolation from the rest of the cluster occurs, it knows exactly where it could
be reassigned preserving the network connectivity.
The Fig. 2 summarizes the steps of the approach proposed in this work:
Fig. 2. Flow-chart showing the steps from the time when a new node joins the network
to the cluster creation
A Self-organization Technique in Wireless Sensor Networks 847
Notation
Given the node N , its backup-clusters list is computed as the list of sensors
in its communication range, but not in its cluster:
backup-clusters(N ) = {n1 , . . . , nk }
In this section, we show the network behavior after an error occurrence, when
our clusterization technique has been used. In order to solve a node crash, we
need to distinguish two phases: the detection and the recovery.
3.3.1 Detection
In this phase, the aim is to understand that a node crashed. A sensor in a
wireless sensor network may shut down because of low battery, or any kind of
fault. Whatever the cause is, we just want to find a way to establish that a
crash occurred. Differently from other proposals, we don’t focus on the crashed
node, but on the one that is affected by the fault. Indeed, we are not interested
in letting the clusterhead know which sensor crashed, but we want to let the
isolated node be aware of its isolation, so that it can be fixed.
In each cluster there is a periodic information exchange: in particular, we
assume a communication scheme such that the clusterhead receives updates by all
the sensors in its cluster at arbitrary time intervals, and it sends back a “received”
answer to all the nodes that have sent updates (both in one or multi hops).
When a fault occurs, if a sensor remains isolated from the rest of the cluster,
it does not receive the answer after sending its updates, and it automatically
detects the error.
848 W. Balzano and S. Stranieri
3.3.2 Recovery
The recovery strategy, following a node crash in a network clusterized with the
algorithm explained above, is made simple by the backup-clusters list. The node
I, that finds itself in isolation from the rest of the cluster, can check its backup-
clusters list and join the cluster identified by the first item of the list, by broad-
casting a message containing its id and the id of the cluster to join. The neigh-
bor(s) receiving this message checks its cluster id and if it matches, the message
is propagated, via multi-hop if needed, to the new clusterhead that adds the new
node to its cluster.
Notice that if the crashed node is a clusterhead, a total reassignment occurs,
meaning that all the sensors in the clusters find themselves in isolation, and they
all apply this recovery technique, reassigning themselves to new clusters. In order
to avoid frequent total reassignments, we introduced in the previous section the
requirement of electing the clusterhead as the sensor with highest energy.
4 Conclusions
In this paper, we address the problem of improving the self-organizing capability
of a wireless sensor network. In particular, we guarantee the network connectivity
even after exceptional errors or faults of a single sensor. The self-organization
property is obtained by exploiting clustering techniques, that allow the network
to minimize the flow of messages, and the communication between nodes, which
constitute the most expensive operation in terms of energy. We want to avoid the
situation where, because of a sensor crash, another sensor cannot communicate
with its cluster anymore. For this reason, provide a clusterization algorithm
letting each sensor keep track of a backup list of clusters, to which it can be
reassigned in case it remains isolated in its cluster.
References
1. Amato, F., Boselli, R., Cesarini, M., Mercorio, F., Mezzanzanica, M., Moscato, V.,
Persia, F., Picariello, A.: Challenge: processing web texts for classifying job offers,
pp. 460–463 (2015). Cited By 16
2. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for E-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). Cited By 18
3. Amato, F., Mazzocca, N., Moscato, F.: Model driven design and evaluation of
security level in orchestrated cloud services. J. Netw. Comput. Appl. 106, 78–89
(2018). Cited By 2
4. Amato, F., Moscato, F.: Model transformations of mapreduce design patterns for
automatic development and verification. J. Parallel Distrib. Comput. 110, 52–59
(2017). Cited By 3
5. Arora, A., Dutta, P., Bapat, S., Kulathumani, V., et al.: A line in the sand: a
wireless sensor network for target detection, classification, and tracking. Comput.
Netw. 46(5), 605–634 (2004)
6. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN-vehicle-2-vehicle elastic network.
Procedia Comput. Sci. 98, 497–502 (2016)
8. Balzano, W., Murano, A., Vitale, F.: WiFACT–wireless fingerprinting automated
continuous training. In: Proceedings of WAINA. IEEE Computer Society (2016)
9. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
10. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: 3PGCIC. Springer (2016)
11. Balzano, W., Del Sorbo, M.R., Stranieri, S.: A logic framework for C2C network
management. In: Proceedings of WAINA. IEEE Computer Society (2016)
12. Balzano, W., Vitale, F.: DiG-Park: a smart parking availability searching method
using V2V/V2I and dgp-class problem. In: Proceedings of WAINA. IEEE Com-
puter Society (2017)
13. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. Int. J. Grid Util. Comput. (2017)
14. Balzano, W., Stranieri, S.: Cooperative localization logic schema in vehicular ad
hoc networks. In: International Conference on NBis, pp. 960–969. Springer (2018)
850 W. Balzano and S. Stranieri
15. Bandyopadhyay, S., Coyle, E.J.: An energy efficient hierarchical clustering algo-
rithm for wireless sensor networks. In: Proceedings of IEEE INFOCOM 2003. IEEE
(2003)
16. Chen, W., Chen, L., Chen, Z.-l., Tu, S.-l.: A realtime dynamic traffic control system
based on wireless sensor network. In: 34th International Conference on Parallel
Processing Workshops. IEEE Computer Society (2005)
17. Dressler, F.: Self-organization in Sensor and Actor Networks. Wiley, Chichester
(2008)
18. Gupta, G., Younis, M.F.: Fault-tolerant clustering of wireless sensor networks. In:
IEEE WCNC 2003, New Orleans, 16–20 March 2003. IEEE (2003)
19. He, T., Krishnamurthy, S., Stankovic, J.A., et al.: Energy-efficient surveillance
system using wireless sensor networks. In: Proceedings of the 2nd International
Conference on Mobile Systems, Applications, and Services, pp. 270–283. ACM
(2004)
20. Krishnamachari, B., Iyengar, S.: Distributed Bayesian algorithms for fault-tolerant
event region detection in wireless sensor networks. IEEE Trans. Comput. 53(3),
241–250 (2004)
21. Krishnan, R., Starobinski, D.: Efficient clustering algorithms for self-organizing
wireless sensor networks. Ad Hoc Netw. 4(1), 36–59 (2006)
22. Liao, Y., Qi, H., Li, W.: Load-balanced clustering algorithm with distributed self-
organization for wireless sensor networks. IEEE Sens. J. 13(5), 1498–1506 (2013)
23. Mainwaring, A., Culler, D., Polastre, J., Szewczyk, R., Anderson, J.: Wireless sen-
sor networks for habitat monitoring. In: Proceedings of the 1st ACM International
Workshop on Wireless Sensor Networks and Applications, pp. 88–97. ACM (2002)
24. Sharma, S., Bansal, R.K., Bansal, S.: Issues and challenges in wireless sensor net-
works. In: International Conference on Machine Intelligence and Research Advance-
ment (ICMIRA), pp. 58–62. IEEE (2013)
25. Sohrabi, K., Gao, J., Ailawadhi, V., Pottie, G.J.: Protocols for self-organization of
a wireless sensor network. IEEE Pers. Commun. 7(5), 16–27 (2000)
26. Zhang, Z., Long, K., Wang, J.: Self-organization paradigms and optimization
approaches for cognitive radio technologies: a survey. IEEE Wirel. Commun. 20(2),
36–42 (2013)
Data Dissemination in Vehicular Ad Hoc
Network: A Model to Improve Network
Congestion
1 Introduction
Imagine a driver, hurrying for an important job meeting, being slow down
because of a road accident along the road he is cruising. What if the driver
had been previously warned about the accident? For sure, he could have chosen
a different path getting to his meeting on time, as well as the road safety would
have been improved. Inter-vehicular communication is thought to solve these
situations. VANETs (Vehicular ad Hoc Networks) are made of vehicles form-
ing a decentralized network, and communicating via On-board units (OBUs).
This kind of networks are based on short-range communication between vehi-
cles: neighbors can communicate directly, while the information propagation is
realized through multi-hop communication.
Given the numerous application fields, such as parking [10,11,16], network
management [7] and classification [2], cluster-oriented communication [6,9], com-
munication optimization [8,13–15], and since the number of vehicles equipped
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 851–859, 2019.
https://doi.org/10.1007/978-3-030-15035-8_83
852 W. Balzano and S. Stranieri
2 Related Work
through Car2Car communication. In this work, Petri Net are used to represent
the behavior of a VANET aimed to reduce redundant computations, so to prevent
network congestion.
Before analyzing the details of the approach proposed in this work, we fix some
basic notations about how to interpret a vehicular network.
Definition 1. A VANET can be seen as a directed graph G = <V, E>, where
the set of vertices V is the set of vehicles, while the set of edges E is such that
the pair (u, v) ∈ E if and only if v is in the communication range of u, in terms
of RSSI (received signal strength indicator).
As we already mentioned, network congestion has been defined in different ways
by the researchers. It is essentially a network condition in which it is no longer
possible to accomplish all the incoming requests. In a vehicular context, network
congestion may happen in situations of high traffic density: for instance, when
a vehicle receives different messages coming from different neighbors, it has to
schedule their transmission according to some metrics. This is what researchers
do by means of congestion control mechanisms, often, as we have mentioned,
by associating priorities to messages. Without loss of generality, in this work we
assume that vehicles move from west to east and, hence, the message propa-
gation flows from east to west. We want to study the network congestion from
another perspective with respect to other researchers. Indeed, we observed that
redundancy in transmission may impact the congestion for the receiving vehicle
and, as a consequence, for the whole network. As Fig. 1 shows, if some event
occurs along the flow direction, the A vehicle would detect it before the others
and would transmit the message to its neighbors in the propagation direction: B
and C. These two vehicles would in turn propagate the information to D, that
would receive the same message twice.
Clearly, the higher the number of nodes is the higher the impact of redundancy
is, since the network gets more congested. Moreover, in a simulation with only 40
nodes in the network, the number of repetitions of the same message overcomes
300, in case of very connected networks. If the network becomes made by hun-
dreds of nodes, the redundancy can seriously impact the VANET safety. This is
a consequence of the fact that, without any operation aimed to redundancy con-
trol, a node receives the same message as many times as the number of different
paths, following the propagation direction, from the source of the message to the
node itself. These observations let evident the need of redundancy elimination
mechanism beyond that congestion control one.
856 W. Balzano and S. Stranieri
simply discards known messages, allowing a clever use of the available resources
that are employed to process and propagate only new messages. This simple
preliminary step could seriously improve the network scalability, since, as we have
seen from the graphics in Fig. 3, redundancy in a high density network can lead to
a very high number of repeated messages. Clearly, congestion control strategies
are still needed: indeed, it may happen that more events occur in areas of the
network which are far from each other, but their propagation converges to the
same node that receives more messages than allowed according to its resources. In
this case, even if redundancy elimination has been previously applied, congestion
control mechanism is required.
4 Conclusions
Network congestion is a big deal in VANET research field. Many congestion
control proposals can be found in literature, but the redundancy contribution
to this phenomenon has not had the same attention. For this reason, in this
work we provide an analysis of the impact of communication redundancy on the
network congestion problem, that highlights the need to manage this problem,
even before the congestion occurs. To this aim, we propose a model, exploiting
the Petri Net modeling language, that explains the behavior of a generic vehicle
of the network when a message is received, and the way the propagation is
handled.
References
1. R-Project: What is R? https://www.r-project.org/about.html
2. Amato, F., Boselli, R., Cesarini, M., Mercorio, F., Mezzanzanica, M., Moscato, V.,
Persia, F., Picariello, A.: Challenge: processing web texts for classifying job offers,
pp. 460–463 (2015). Cited By 16
858 W. Balzano and S. Stranieri
3. Amato, F., Colace, F., Greco, L., Moscato, V., Picariello, A.: Semantic processing
of multimedia data for E-government applications. J. Vis. Lang. Comput. 32, 35–41
(2016). Cited By 18
4. Amato, F., Mazzocca, N., Moscato, F.: Model driven design and evaluation of
security level in orchestrated cloud services. J. Netw. Comput. Appl. 106, 78–89
(2018). Cited By 2
5. Amato, F., Moscato, F.: Model transformations of MapReduce design patterns for
automatic development and verification. J. Parallel Distrib. Comput. 110, 52–59
(2017). Cited By 3
6. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN-vehicle-2-vehicle elastic network.
Procedia Comput. Sci. 98, 497–502 (2016)
8. Balzano, W., Murano, A., Vitale, F.: WiFACT-wireless fingerprinting automated
continuous training. In: Proceedings of WAINA. IEEE Computer Society (2016)
9. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: 3PGCIC. Springer (2016)
10. Balzano, W., Vitale, F.: DiG-Park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: Proceedings of WAINA. IEEE Com-
puter Society (2017)
11. Balzano, W., Barbieri, V., Riccardi, G.: Smart priority park framework based on
DDGP3. In: 32nd International Conference on Advanced Information Networking
and Applications Workshops (WAINA), pp. 674–680. IEEE (2018)
12. Balzano, W., Rosaria, M., Sorbo, D., Stranieri, S.: A logic framework for C2C
network management. In: 30th International Conference on Advanced Information
Networking and Applications Workshops (WAINA), pp. 52–57. IEEE (2016)
13. Balzano, W., Formisano, M., Gaudino, L.: WiFiNS: a smart method to improve
positioning systems combining WiFi and INS techniques. In: International Con-
ference on Intelligent Interactive Multimedia Systems and Services, pp. 220–231.
Springer (2017)
14. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. In: The International Conference on Data Compression,
Communication, Processing and Security, CCPS 2016 (2016)
15. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
16. Balzano, W., Stranieri, S.: LoDGP: a framework for support traffic information
systems based on logic paradigm. In: International Conference on P2P, Parallel,
Grid, Cloud and Internet Computing, pp. 700–708. Springer (2017)
17. Balzano, W., Stranieri, S.: Cooperative localization logic schema in vehicular ad
hoc networks. In: International Conference on Network-Based Information Sys-
tems, pp. 960–969. Springer (2018)
18. Balzano, W., Stranieri, S.: A logic user-based algorithm to improve node distribu-
tion in wireless sensor network. In: VLSS 2018. Springer (2018)
19. Bansal, G., Cheng, B., Rostami, A., Sjoberg, K., Kenney, J.B., Gruteser, M.:
Comparing LIMERIC and DCC approaches for VANET channel congestion con-
trol. In: IEEE 6th International Symposium on Wireless Vehicular Communica-
tions (WiVeC), pp. 1–7. IEEE (2014)
20. Bouassida, M.S., Shawky, M.: On the congestion control within VANET. In: Wire-
less Days, WD 2008. 1st IFIP, pp. 1–5. IEEE (2008)
21. Di Febbraro, A., Giglio, D., Sacco, N.: Urban traffic control structure based on
hybrid petri nets. IEEE Trans. Intell. Transp. Syst. 5(4), 224–237 (2004)
Data Dissemination in Vehicular Ad Hoc Network 859
22. Djahel, S., Ghamri-Doudane, Y.: A robust congestion control scheme for fast and
reliable dissemination of safety messages in VANETs. In: Wireless Communications
and Networking Conference (WCNC), pp. 2264–2269. IEEE (2012)
23. Fowler, H.J., Leland, W.E., Bellcore, B.: Local area network traffic characteristics,
with implications for broadband network congestion management. IEEE J. Sel.
Areas Commun. 9(7), 1139–1149 (1991)
24. Julvez, J.J., Boel, R.K.: A continuous petri net approach for model predictive
control of traffic systems. IEEE Trans. Syst. Man Cybern. Part A Syst. Hum.
40(4), 686–697 (2010)
25. Mughal, B.M., Wagan, A.A., Hasbullah, H.: Efficient congestion control in VANET
for safety messaging. In: International Symposium in Information Technology
(ITSim), vol. 2, pp. 654–659. IEEE (2010)
26. Murano, A., Perelli, G., Rubin, S.: Multi-agent path planning in known dynamic
environments. In: PRIMA 2015. LNCS, vol. 9387, pp. 218–231. Springer (2015)
27. Murata, T.: Petri nets: properties, analysis and applications. Proc. IEEE 77(4),
541–580 (1989)
28. Nadeem, T., Shankar, P., Iftode, L.: A comparative study of data dissemination
models for VANETs. In: Third Annual International Conference on Mobile and
Ubiquitous Systems: Networking & Services, pp. 1–10. IEEE (2006)
29. Rubin, S., Zuleger, F., Murano, A., Aminof, B.: Verification of asynchronous
mobile-robots in partially-known environments. In: PRIMA 2015. LNCS, vol. 9387,
pp. 185–200. Springer (2015)
30. Sepulcre, M., Gozalvez, J., Härri, J., Hartenstein, H.: Contextual communications
congestion control for cooperative vehicular networks. IEEE Trans. Wirel. Com-
mun. 10(2), 385–389 (2011)
31. Tielert, T., Jiang, D., Chen, Q., Delgrossi, L., Hartenstein, H.: Design methodol-
ogy and evaluation of rate adaptation based congestion control for vehicle safety
communications. In: Vehicular Networking Conference (VNC), pp. 116–123. IEEE
(2011)
32. Welzl, M.: Network Congestion Control: Managing Internet Traffic. Wiley, Chich-
ester (2005)
33. Yang, X., Liu, L., Vaidya, N.H., Zhao, F.: A vehicle-to-vehicle communica-
tion protocol for cooperative collision warning. In: The First Annual Interna-
tional Conference on Mobile and Ubiquitous Systems: Networking and Services,
MOBIQUITOUS 2004, pp. 114–123. IEEE (2004)
34. Yousefi, S., Mousavi, M.S., Fathy, M.: Vehicular ad hoc networks (VANETs): chal-
lenges and perspectives. In: Proceedings of the 2006 6th International Conference
on ITS Telecommunications, pp. 761–766. IEEE (2006)
ManDri: A New Proposal of Manus VR
Facility Integration in Everyday Car
Driving
Abstract. The purpose of this paper is to analyze all the possible uses
of the Manus VR in everyday driving, to make it easier and safer. With
these technological gloves, equipped with eleven sensors for each hand,
able to precisely pick up every movement of the hand and fingers, it
is possible to associate various gestures to specific actions. In particu-
lar, they allow to turn, accelerate, brake, activate the car turn signals,
the air conditioning, even alarm, everything just using simple gestures.
But ManDri is not only framed in today’s guide: indeed, a glove, being
equipped with two IMUs (inertial measurement units), can be integrated
with GPS (global position system). In fact, progresses in these years allow
the production of IMU-enabled GPS devices, that allow a GPS receiver
to work even when GPS-signals are unavailable, such as inside buildings,
tunnels, or places characterized by electronic interference. Moreover, it
is also possible to create the ManDri app for all the devices so that, after
connecting them to the gloves, dedicated accounts can be created for
chosen people allowed to use them, making easier the control, the man-
agement, and the gloves gestures customization. This device unavoidably
becomes essential and difficult to replace after trying it. These positive
premises make us understand how advantageous the choice to use these
gloves in everyday driving to improve it, make it safer, and undoubtedly
facilitate it.
1 Introduction
2 Related Work
VANET is a growing research field aimed to improve road safety and traffic
conditions [2,5], and path tracking techniques [3]. Various proposals have been
done to guarantee a “virtual guide”, or a connection man-car with some gestures,
but never exploiting Manus VR, meaning never with a device that guarantees
such precision. For example in [23], they presents a prototype in-car system
that allows car features (like turning lights and windows) to be controlled by
combinations of speech, gaze, and micro-gestures. They propose an interaction
concept man-car where “Micro Gestures” are directly performed with the hands
on the steering wheel, so it can be considered hands-free. For the recognition of
finger movements they use a Leap Motion controller, which is placed at the center
of the steering wheel. This is a small USB device which uses two monochrome
IR cameras at 200 Hz wide-angle (up to 150◦ ) and three infrared LEDs, that
permit to observe a certain area for a distance of about 1 m.
Similar to the Leap Motion mentioned before, authors in [14] introduce the
design method of intelligent car based on Wi-Fi video capture. The intelligent
car designed in this paper combines Wi-Fi technology and image processing
technology, and is based on Wi-Fi video capture and OpenCV gesture control
ManDri: A New Proposal of Manus VR Facility Integration 863
[19]. The car can carry out on-site video capture by the USB camera installed
on a router wireless, then transfers through it the collected video streaming to
the PC side, in order to achieve on-site surveillance.
The main protagonist of gesture recognition is the Kinect [15,24,29]. In [26],
for example, they have used Microsoft’s Kinect sensor to detect the human ges-
tures in a driver game. This device primarily has three sensors: IR emitter, IR
detector and RGB camera. Based on these sensors along with software develop-
ment kit (SDK), a human body is approximated using 20 body joint coordinates
in 3-dimensional space. However, for the implementation of the virtual reality
problem of car racing game, only the information on 9 joints are necessary. Using
this reduced set of sensors reduces the computation time, which is the prime con-
cern for real-time systems. The working range for Kinect sensor is from 1.2 m to
3.5 m, which makes this device perfectly acceptable for application in car racing
games, but nothing more.
Instead, in [28], they use a mm-wave radar sensor. Radar is advantageous
because it is not affected by light or sound in the environment; it (i ) can be
embedded in devices, (ii ) has very precise resolution, (iii ) offers real time recog-
nition, and (iv ) does not require recording of an image of the user. Their problem
is when they are placed too close to detectable objects, such as a gear shift. In
these cases, it often recognizes false positives, or does not recognize gestures.
In [21], they present an original contactless human machine interface for
driving car. The proposed framework is based on an image sent by a simple
camera device, which is then processed by various computer vision algorithms.
These algorithms allow the isolation of the user’s hand on the camera frame and
the translation of its movements into orders, that are sent to the computer in a
real time process. The optimization of the implemented algorithms on graphics
processing unit leads to real time interaction between the user, the computer,
and the machine. The user can easily modify or create the interfaces displayed
by the proposed framework to fit his personnel needs. Their framework works
in two steps. The first one consists in the initialization: an XML file is read,
containing the configuration to use; the second step consists in the execution of
the previous one.
In [25], they created a capacitive sensing method using electrode as sensing
input. The electrode acts as detector to retrieve the field variations at different
position of hand, in order to measure the origin of electric field distortion from the
variation in signals received due to the conductivity of the human body intercept
with capacitor itself. Then this information is used to calculate the position
of hand, to track movement of hands, and to classify gestures. This system
consists of three aluminum foil board, structured it in cube, from with each plate
acting as X, Y, Z coordinates. The cube structure is conductive so that it will
store capacitive charge with current and voltage supplied it, so the movement
of hand within the 3 plates of capacitor will act as a gesture. These gestures
will discharge the capacitor. The discharging time of each plate is different and
varies accordingly. The discharging values of each plate change according to the
situation: for instance, they very if capacitor is placed on wooden material, or on
864 W. Balzano et al.
3 ManDri
With a simple movement of the hand with the thumb raised to the left you can
decide whether to take the road on left (Fig. 1a). The same specular movement,
instead, will allow you to take the right one (Fig. 1b). With a movement of the
hand with the index raised forward it will be possible to start or accelerate the
car, depending on whether it is stationary or moving (Fig. 1c). It will be the
CheckFeasibility algorithm to manage this logic, in fact in the algorithm there
is a check on the current state of the car, and based on it, it will be possible to
understand the action that the car must perform, in this case start or accelerate.
Instead a backward movement with the thumb will make it decelerate, stop or
reverse gear (Fig. 1d). By doing the pinch with index and thumb, as shown in
Fig. 1e, it will be possible to change the command mode, moving from the driving
gestures to the gestures for the GPS.
We can imagine how many gestures are possible, and how many components
we can apply this technology: turn signals, seats, air conditioning, lights, anti-
theft, in short, to every component of the car. The usefulness of this technology
takes its maximum level in dangerous situations. In these situations, the human
866 W. Balzano et al.
Fig. 1. The turn left (a), right (b), start/accelerate (c), stop/decelerate/reverse (d),
and change mode (e) gesture.
instinct urges us to put our hands forward to protect ourselves, or make a sudden
turn, then a sudden movement, in the same way in the guide if a sudden obstacle
appears before us, like a dog, a landslide; instinct would push us to protect
us, reading and processing these gestures we could make sure to stop the car
immediately, certainly stopping it before it would stop if the driver should press
the brake pedal, saving the life of the passenger and the life of those around us.
This idea can perfectly be combined with [8].
The main character of the system is Manus VR. These are gloves for high-end
touch data that have been created to capture human hands motion to use them
as a perfect input device. They are controllers capable of capturing and repro-
ducing, respectively, the position and the movements of the hands and fingers,
by means of advanced trackers and sensors. Their shape is different from a hypo-
thetical controller, because these are real technological gloves. One glove contains
two inertial measurement units, one located on the top of your hand and one
on your thumb which measures it moving separately from the other fingers.
An IMU is an electronic device that can calculate and reports an angular rate,
body’s specific force, and sometimes the magnetic field surrounding the body,
ManDri: A New Proposal of Manus VR Facility Integration 867
5 Experimental Results
the gestures implemented in the HGR (Hand Gesture Recognition) System [11].
Results are shown in Table 1.
As we can see, results for almost all the gestures are almost perfect: indeed,
the lowest success rate is 97%. The lowest success rate is obtained with the pinch
gesture, that is recognized less than the others.
References
1. Ahmed, A., Loo, C., Obo, T.: Neuro-fuzzy model with subtractive clustering opti-
mization for arm gesture recognition by angular representation of kinect data. In:
6th ICIEV-ISCMHT, pp. 1–6. IEEE (2017)
2. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
3. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. Int. J. Grid Util. Comput. (2017)
4. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
5. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: 3PGCIC. Springer (2016)
6. Balzano, W., Stranieri, S.: Cooperative localization logic schema in vehicular ad
hoc networks. In: International Conference on NBis, pp. 960–969. Springer (2018)
7. Balzano, W., Stranieri, S.: A logic range-free algorithm for localization in wireless
sensor networks. In: The 24th International DMSVIVA (2018)
8. Balzano, W., Vitale, F.: RADS: a smart road anomalies detection system using
vehicle-2-vehicle network and cluster features
9. Balzano, W., Vitale, F.: FiDGP: a smart fingerprinting radiomap refinement
method based on distance-geometry problem. In: International Conference on
NBis, pp. 970–978. Springer (2018)
ManDri: A New Proposal of Manus VR Facility Integration 869
10. Balzano, W., Vitalem F.: GER-EN–GNSS error reduction using an elastic network
based on V2V and LIDAR. In: International Symposium on Cyberspace Safety
and Security, pp. 124–131. Springer (2018)
11. Benmoussa, M., Mahmoudi, A.: Machine learning for hand gesture recognition
using bag-of-words. In: ISCV, pp. 1–7. IEEE (2018)
12. Bhuiyan, R., Tushar, A., Ashiquzzaman, A., Shin, J., Islam, M.: Reduction of
gesture feature dimension for improving the hand gesture recognition performance
of numerical sign language. In: 20th International Conference of ICCIT, pp. 1–6.
IEEE (2017)
13. Chiang, T., Fan, C.: 3D depth information based 2D low-complexity hand posture
and gesture recognition design for human computer interactions. In: 3rd ICCCS,
pp. 233–238. IEEE (2018)
14. Dai, G., Wang, P.: Design of intelligent car based on WiFi video capture and
OpenCV gesture control. In: CAC, pp. 4103–4107. IEEE (2017)
15. Dai, G., Yu, L., Huang, J.: Dynamic and interactive gesture recognition algorithm
based on kinect. In: CCDC, pp. 3479–3484. IEEE (2017)
16. Dehankar, A., Jain, S., Thakare, V.: Performance analysis of RTEPI method for
real time hand gesture recognition. In: ICISS, pp. 1031–1036. IEEE (2017)
17. Gajjar, V., Mavani, V., Gurnani, A.: Hand gesture real time paint tool-box:
machine learning approach. In: ICPCSI, pp. 856–860. IEEE (2017)
18. Hussain, S., Saxena, R., Han, X., Khan, J., Shin, H.: Hand gesture recognition
using deep learning. In: ISOCC, pp. 48–49. IEEE (2017)
19. Khan, M., Mishra, K., Qadeer, M.: Gesture recognition using OpenCV. In: 7th
CSNT, pp. 167–171. IEEE (2017)
20. Lei, T., Jia, X., Zhang, Y., Zhang, Y., Su, X., Liu, S.: Holoscopic 3D micro-gesture
recognition based on fast preprocessing and deep learning techniques. In: Auto-
matic Face & Gesture Recognition (FG 2018), pp. 795–801. IEEE (2018)
21. Magoulès, F., Zou, Q.: GPU accelerated contactless human machine interface for
driving car. In: 16th International Symposium on DCABES, pp. 7–10. IEEE (2017)
22. Maharani, D., Fakhrurroja, H., Machbub, C., et al.: Hand gesture recognition using
k-means clustering and support vector machine. In: ISCAIE 2018, pp. 1–6. IEEE
(2018)
23. Neßelrath, R., Moniri, M., Feld, M.: Combining speech, gaze, and micro-gestures
for the multimodal control of in-car functions. In: 12th International Conference
on IE, pp. 190–193. IEEE (2016)
24. Plouffe, G., Cretu, A.: Static and dynamic hand gesture recognition in depth data
using dynamic time warping. IEEE Trans. Instrum. Meas. 65(2), 305–316 (2016)
25. Rau, M., Agarkar, B.: 3D GRS for car infotainment. In: International Conference
on ICICET, pp. 1–4. IEEE (2018)
26. Saha, S., Lahiri, R., Konar, A., Ralescu, A., Nagar, A.: Implementation of gesture
driven virtual reality for car racing game using back propagation neural network.
In: SSCI, pp. 1–8. IEEE (2017)
27. Sasaki, A., Hashimoto, H.: High degree-of-freedom hand model driven by lower
degree-of-freedom input. In: SII, pp. 722–727. IEEE (2017)
28. Smith, K., Csech, C., Murdoch, D., Shaker, G.: Gesture recognition using mm-wave
sensor for human-car interface. IEEE Sens. Lett. 2(2), 1–4 (2018)
29. Varshini, M., Vidhyapathi, C.: Dynamic fingure gesture recognition using kinect.
In: ICACCCT, pp. 212–216. IEEE (2016)
30. Wang, K., Zhao, R., Ji, Q.: Human computer interaction with head pose, eye
gaze and body gestures. In: Automatic Face & Gesture Recognition (FG 2018),
pp. 789–789. IEEE (2018)
A Smart Compact Traffic Network Vision
Based on Wave Representation
1 Introduction
Wireless Sensor Networks (WSNs) are made of low-cost and low-power sensors
able to communicate and perform distributed tasks, often by self-organizing
into clusters. These networks can be largely employed for different purposes,
such as sensing, event detection, localization [8,9,11,21]. The WSN features of
congestion control, self-configuration, and energy awareness let their use keep
growing. In particular, congestion in WSN constitutes a concrete challenging
issue, since it determines a network performance decrease, as a consequence of
an amount of requests which is higher than the ones it can be actually satisfied.
As pointed out in [3], the increasing interest in the last years is due to the
fact that rapid deployment and fault tolerance characteristics of sensor networks
make them a promising sensing technique for military purposes. Moreover, the
sensor ability of tracking the movement of small animals, as well as monitoring
the surrounding ambient, makes them suitable for environmental aims.
An interesting application of WSN is in the vehicular context [22]: indeed, a
new emerging field is the vehicular ad hoc network, where the vehicles play the
role of sensors in a WSN [4]. Due to the increasing number of road accidents,
the need for a stable communication system between vehicles is growing. Cars
can broadcast data and information about exceptional events occurrence via
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 870–879, 2019.
https://doi.org/10.1007/978-3-030-15035-8_85
A Smart Compact Traffic Network Vision Based on Wave Representation 871
wireless media, letting the rest of the vehicles be aware of the condition inside
the network [16]. This mechanism has a significant impact not only on safety,
but also in traffic management. Due to the importance of such a network, in this
work we provide a smart VANET representation by including in a single compact
wave a lot of useful information about node congestion, traffic, and network
clusterization. Such a wave is obtained by computing the value of congestion,
named congestion factor, of any vehicle in the network, so that highly jammed
traffic areas can be immediately detected by identify the highest peaks of the
wave. Essentially, we propose a way to obtain the highest amount of information
about the network, with as few data as possible.
Outline of the Paper
The rest of the paper is organized as follows. In Sect. 2, the state of the art
is analyzed. In Sect. 3, our novel network representation is provided. Here we
distinguish the phase in which the signal is built from the one in which the
signal is analyzed and understood. In this section we also provide some examples
of use, just for the purpose of illustration. Finally, Sect. 4 gives the conclusions
and some hints for future developments of this paper, working mainly on the
obtained wave.
2 Related Work
3.1 Preliminaries
As follows, we provide some basic concepts used in the rest of the paper. As
these are common definitions, an expert reader can skip this part.
We consider a traffic network as a undirected graph G = (V, E) in which the
set of nodes V corresponds to the vehicles in the network, with |V | = n, and the
set of edges E corresponds to the links, if any, between the nodes, with |E| = m.
We define R as the distance coefficient, i.e. the maximum value within which
two vehicles can communicate with each other. Such a value is calculated refer-
ring to the measurement of the power present in a received radio signal, i.e. the
received signal strength indicator (RSSI). The RSSI values are measured in dBm
and have typical negative values ranging between 0 dBm (excellent signal) and
−110 dBm (extremely poor signal) [20]. In particular, it is intuitive to say that
A Smart Compact Traffic Network Vision Based on Wave Representation 873
increasing the distance, the RSSI decreases. Moreover, given two nodes u, v ∈ V,
we define d(u, v) ∈ [0, R] the euclidean distance between u and v. We can state
that an edge (u, v) ∈ E exists iff d(u, v) ≤ R, and we indicate with weight(u, v)
its weight.
Since in a vehicular network d(i, j) = d(j, i) (with i = j), we directly provide a
normalized graph, by computing the average of these two distances and by assign-
ing this value to the edge linking i and j, so that the corresponding adjacency
matrix is symmetric. Finally, for each v ∈ V define neig(v) = {v |(v, v ) ∈ E}
and degree(v) = |neig(v)| as the set of v-neighbors and its number, respectively.
We can observe that the congestion degree of a node does not depend only
on the number of its neighbors, but mostly on their congestion. Thus, we need
to take into account the congestion of each neighbor of the considered node, in
874 W. Balzano et al.
order to avoid the situation as the one shown in Fig. 1, in which the nodes A and
B have the same number of neighbors, but the congestion of such neighbors is
different. Indeed, in the left side we have a higher congestion level with respect
to the right side, but with the previous formula A and B would have had the
same congestion factor.
For this reason, we introduce a weighted ideal congestion:
weighted ideal(v) = (degree(u) − 1) ∗ ideal(v) (4)
u∈neig(v)
Notice that when we compute the degree of the neighbors of a given node v,
we decrease it by 1 in order not to consider v again. Once the congestion factors
are computed, they are normalized, dividing them by the maximum congestion
factor of the considered network.
The network representation we want to build is f (V ), with f as defined in
Eq. 1. A key point in the wave representation is the order in which the nodes are
placed on the x axis. Our goal is to obtain a signal that immediately highlights
the most congested areas and which points belong to them. For this reason,
we propose a cluster-oriented visit of the network that, starting from a random
point, continues the visit of neighbors, putting them in the same connected area,
as long as a certain distance is not overcome. With “connected area” we mean a
set of related nodes. It is important to notice that this is not a new clusterization
technique, but an alternative way of visiting a network by detecting areas of
nodes connected within a certain distance. To this aim, we introduce a tolerance
such that, given a node v and its neighbor u, if d(v, u) ≤ R − , u is in the same
connected area as v. It is necessary to notice that we distinguish the congestion
factor of isolated nodes by the one of nodes having only edges with weight ≥R−,
by preserving them in the network. Once the connected areas are obtained, the
nodes are placed on the x axis in such a way that the ones belonging to the same
connected area are contiguous. In particular, for each connected area we put
the elements belonging to it in increasing order with respect to the congestion
factor. This choice makes easier the detection of connected areas looking at the
signal.
Fig. 2. Random generation of a network with n = 35, sd = 10, R = 10, = 0.3 (a)
and the corresponding wave representation (b)
not hold in general. Hence, this is a necessary but not a sufficient condition for
the starting of a new area. This means that there could be changes of area
hidden by the signal, when the highest congestion factor of the first connected
area is smaller than the lowest one of the second connected area, introducing
false negatives. As shown in Fig. 2(b), which is the signal corresponding to the
graph Fig. 2(a), it is intuitive to observe that the most congested nodes are 24
and 29. Another information easy to deduce is about the areas of the network
made by single nodes. Indeed, they are identified by the points whose congestion
factor coincides with the dashed line, i.e. 3, 11, 33, and 35. In order to detect
the remaining connected areas, we need to retrieve the high-low transitions,
corresponding in this network to the points 9, 29, 26, and 27. By analyzing the
network in Fig. 2(a), we would have expected a change of area between the nodes
4, 13 and 8, 17, 22, that is hidden by the signal because of false negatives. The
same happens between the nodes 22 and 7.
The signal obtained as shown in the examples, is determined by visiting the
nodes of the network starting from the one having the smallest x-coordinate.
By changing the starting point, clearly the congestion factor of each node
stays the same, but the resulting wave can be a permutation of the peaks in
the current signal. This could be a limitation for comparisons between waves.
876 W. Balzano et al.
For this reason, we introduce a canonical form of the signal, obtained by chang-
ing the order of the nodes on the x axis: not only the nodes are ordered in
increasing order of congestion factor inside each connected area, but also each
connected area is ordered in decreasing order of maximum congestion factor on
the x axis. Through this normalization, we obtain a signal having the connected
area with the highest congested node on the left side, and the single nodes on
the right side.
With this approach, false negatives are also reduced. Indeed, it is no longer
possible that the highest congested factor of a previous connected area is smaller
than the lowest factor of the next area, since they are ordered, but false negatives
can still occur, as formally reported in the following theorem:
Theorem 1. Given two successive connected areas a1 and a2 , let hi and si
(with i ∈ [1, 2]) be the highest and lowest congestion factors respectively for the
corresponding area, then:
Fig. 3. Random generation of a network with n = 35, sd = 10, R = 10, = 0.2 (a)
and the corresponding wave representation (b).
A Smart Compact Traffic Network Vision Based on Wave Representation 877
Fig. 4. Random generation of a network with n = 35, sd = 8, R = 10, = 0.2 (a) and
the corresponding wave representation (b).
Proof. Let us assume the premise true, thus we have a false negative. The
following inequalities hold:
⎧
⎨ h1 ≥ h2 by construction
⎪
h1 ≤ s2 by def inition of f alse negative
⎪
⎩
s2 ≤ h2 trivially
The example in Fig. 3(b) shows the signal construction following the random
generation of the corresponding network in Fig. 3(a). As expected, the peaks
decreases from left to right (with 21 the most congested vehicle), each high-low
transition indicates a change of connected area, and the nodes whose values
correspond with the dashed line are isolated vehicles (2, 10, 19, 35).
By slightly changing the standard deviation value, we generate a more con-
nected network such as the one shown in Fig. 4(a). Analyzing the corresponding
signal in Fig. 4(b), we can deduce that the most congested node is 18, there are
three isolated nodes (4, 34, 10), and high-low transitions occur correspondingly
to the nodes 18, 3, 15, 5, 28 and 34 identifying the end of a connected area.
Other examples my be obtained by modifying the input parameters: the
result is a network with a different distribution of nodes. By decreasing the
maximum distance allowed between any pair of nodes, we automatically make
the radio signal less powerful. As a consequence, the visibility between nodes is
878 W. Balzano et al.
References
1. Statista: IoT connected devices installed base worldwide from 2015 to 2025. www.
statista.com/statistics/471264/iot-number-of-connected-devices-worldwide/
2. Statista: Most popular mapping apps in the U.S. 2018, by audience. www.statista.
com/statistics/865413/most-popular-us-mapping-apps-ranked-by-audience/
3. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor net-
works: a survey. Comput. Netw. 38(4), 393–422 (2002)
4. Al-Sultan, S., Al-Doori, M., Al-Bayatti, A., Zedan, H.: A comprehensive survey on
vehicular ad hoc network. J. Netw. Comput. Appl. 37, 380–392 (2014)
5. Balzano, W., Barbieri, V., Riccardi, G.: Smart priority park framework based on
DDGP3. In: WAINA 2018. IEEE (2018)
6. Balzano, W., Murano, A., Stranieri, S.: Logic-based clustering approach for man-
agement and improvement of VANETs. J. High Speed Netw. 23(3), 225–236 (2017)
7. Balzano, W., Murano, A., Vitale, F.: V2V-EN–vehicle-2-vehicle elastic network.
Procedia Comput. Sci. 98, 497–502 (2016)
8. Balzano, W., Murano, A., Vitale, F.: WiFACT–wireless fingerprinting automated
continuous training. In: Proceedings of WAINA 2016. IEEE Computer Society
(2016)
9. Balzano, W., Murano, A., Vitale, F.: EENET: energy efficient detection of network
changes using a wireless sensor network. In: Proceedings of CISIS 2017. Springer
(2017)
A Smart Compact Traffic Network Vision Based on Wave Representation 879
10. Balzano, W., Murano, A., Vitale, F.: Hypaco–a new model for hybrid paths com-
pression of geodetic tracks. Int. J. Grid Util. Comput. (2017)
11. Balzano, W., Murano, A., Vitale, F.: SNOT-WiFi: sensor network-optimized train-
ing for wireless fingerprinting. J. High Speed Netw. 24(1), 79–87 (2018)
12. Balzano, W., Del Sorbo, M.R., Murano, A., Stranieri, S.: A logic-based clustering
approach for cooperative traffic control systems. In: Proceedings of 3PGCIC 2016.
Springer (2016)
13. Balzano, W., Del Sorbo, M.R., Stranieri, S.: A logic framework for C2C network
management. In: Proceedings of WAINA 2016. IEEE Computer Society (2016)
14. Balzano, W., Vitale, F.: DiG-Park: a smart parking availability searching method
using V2V/V2I and DGP-class problem. In: Proceedings of WAINA 2017. IEEE
Computer Society (2017)
15. Boukerche, A., Oliveira, H., Nakamura, E.F., Loureiro, A.: Vehicular ad hoc net-
works: a new challenge for localization-based systems. Comput. Commun. 31(12),
2838–2849 (2008)
16. Cunha, F., Villas, L., Boukerche, A., Maia, G., Viana, A., Mini, R., Loureiro, A.:
Data communication in VANETs: protocols, applications and challenges. Ad Hoc
Netw. 44, 90–103 (2016)
17. Li, F., Wang, Y.: Routing in vehicular ad hoc networks: a survey. IEEE Veh.
Technol. Mag. 2(2), 12–22 (2007)
18. Lochert, C., Hartenstein, H., Tian, J., Fussler, H., Hermann, D., Mauve, M.: A
routing strategy for vehicular ad hoc networks in city environments. In: Intelligent
Vehicles Symposium. IEEE (2003)
19. Naumov, V., Gross, T.R.: Connectivity-Aware Routing (CAR) in vehicular ad-hoc
networks. In: Proceedings of INFOCOM 2007. IEEE (2007)
20. Parameswaran, A.T., Husain, M.I., Upadhyaya, S., et al.: Is RSSI a reliable param-
eter in sensor localization algorithms: an experimental study. In: Proceedings of
F2DA 2009. IEEE (2009)
21. Raghavendra, C.S., Sivalingam, K.M., Znati, T.: Wireless Sensor Networks.
Springer, London (2006)
22. Saha, A.K., Johnson, D.B.: Modeling mobility for vehicular ad hoc networks. In:
Proceedings of VANET 2004. ACM (2004)
23. Singh, A., Kumar, M., Rishi, R., Madan, D.K.: A relative study of MANET and
VANET: its applications, broadcasting approaches and challenging issues. In: Pro-
ceedings of CCSIT 2011. Springer (2011)
24. Vodopivec, S., Bešter, J., Kos, A.: A survey on clustering algorithms for vehicular
ad-hoc networks. In: Proceedings of TSP 2012. IEEE (2012)
25. Zeadally, S., Hunt, R., Chen, Y., Irwin, A., Hassan, A.: Vehicular Ad Hoc Net-
works (VANETs): status, results, and challenges. Telecommun. Syst. 50(4), 217–
241 (2012)
26. Zhang, Z., Boukerche, A., Pazzi, R.: A novel multi-hop clustering scheme for vehic-
ular ad-hoc networks. In: Proceedings of MOBIWAC 2011. ACM (2011)
Big Data Analytics for Traceability
in Food Supply Chain
1 Introduction
More and more, our daily actions leave behind us a digital trail: our consumption
habits, our web searches, videos uploaded or viewed online, opinions shared in
social networks, call logs, even car or our own bodies movements. In fact, the
use of any electronic devices generates information, both direct and indirect, on
people and their interactions that remain recorded and can go to create huge-
sized datasets; about two and a half million (2,500, 000, 000, 000, 000, 000) of
bytes created every day in the world are estimated. This huge availability of data
in real time, for those interested in studying human behaviour, is comparable to
a gold mine to be exploited in some way.
With the current availability of data, it is possible to study various phe-
nomena with a considerably greater amount of information to be analysed and
available at lower costs and times [1,2]. With Big Data we can not only use large
databases to validate a hypothesis already formulated, but also give voice to the
data themselves, making us not only enlighten them, but also get inspired by
them.
In this work we show an application and some examples of the use of Big
Data technologies for traceability in food supply chain domain. The paper is
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 880–884, 2019.
https://doi.org/10.1007/978-3-030-15035-8_86
Big Data Analytics for Traceability in Food Supply Chain 881
structured as follows: Sect. 2 describes the Data model and the architecture of
our system; while Sect. 3 remarks the conclusions and the goals achieved.
2 System Description
2.1 Data Model
The data model relies on the concept of item FI that is in turn related to a given
food item, it is inspired by the data model reported in [3]. In a simplified way, we
consider an ontology O = (V, E) as a network of concepts belonging to a given
domain of interest, where a node v ∈ V represents a “concept” and an edge
e ∈ E represents a relationship between two concepts. We define an annotation
schema and a semantic annotation [4] for a food item as follows.
In other words, the attributes A1 , ..., An are ontological attributes and corre-
spond to concepts that are relevant for the specific domain(s) being modelled. In
turn, non-ontological attributes B1 , ..., Bm can contain other useful information,
such as measures returned by the sensors attached to food items, or multime-
dia objects (e.g., audio, video, images, text, and 3D models) characterized by
low-level features and metadata. In particular, we can adopt both “literals” or a
set of Uniform Resource Identifiers (URIs) that allow access to related informa-
tion according to the Linked Data/Linked Open Data (LD/LOD) paradigms, as
values of the annotation attributes.
2.2 Architecture
with the above-described data model and managed also exploiting the LD/LOD
paradigm. In addition, specific semantics to be attached to the data is provided
using the annotation schemes, including ontologies, vocabularies, taxonomies,
and so on, related to the food traceability domain. The Knowledge Base lever-
ages different advanced data management technologies (e.g., Distributed File
Systems, NoSQL (non relational SQL), and relational systems) and provides a
set of basic APIs to read/write data.
The data management layer is further enriched with two additional com-
ponents: a Geographic Information System (GIS) and a Context Management
System (CMS) [9–11]. The GIS manages the maps of the environments offering
a set of primitives to capture, store, manipulate, retrieve, analyse, and present
all types of spatial or geographical data; the CMS performs data tailoring by
exploiting context information and generates a set of geo-referenced Contextual
Areas, possibly enabling a set of services on such data views [12–15]. As a basis
for the application layer, our system provides a set of services that can be invoked
by user applications:
3 Conclusions
In this article, we showed a model and a system architecture finalized to use Big
Data technologies for traceability in food supply chain domain. In particular, we
described a scalable prototype for the management and context-driven browsing
of food production environments. The system is characterized by several features
that are typical of the modern big data systems and can be used to monitor data
about food supply chain.
References
1. Cilardo, A., Durante, P., Lofiego, C., Mazzeo, A.: Early prediction of hardware
complexity in HLL-to-HDL translation. In: 2010 International Conference on Field
Programmable Logic and Applications (FPL), pp. 483–488. IEEE (2010)
2. Cilardo, A.: Exploring the potential of threshold logic for cryptography-related
operations. IEEE Trans. Comput. 60(4), 452–462 (2011)
884 A. Amato et al.
3. Amato, F., Moscato, V., Picariello, A., Colace, F., De Santo, M., Schreiber, F.A.,
Tanca, L.: Big data meets digital cultural heritage: design and implementation
of scrabs, a smart context-aware browsing assistant for cultural environments.
J. Comput. Cult. Herit. (JOCCH) 10(1), 6 (2017)
4. Amato, F., Mazzeo, A., Penta, A., Picariello, A.: Building RDF ontologies from
semi-structured legal documents. In: 2008 International Conference on Complex,
Intelligent and Software Intensive Systems, CISIS 2008, pp. 997–1002. IEEE (2008)
5. Philip Chen, C.L., Zhang, C.-Y.: Data-intensive applications, challenges, tech-
niques and technologies: a survey on big data. Inf. Sci. 275, 314–347 (2014)
6. Cugola, G., Margara, A.: Complex event processing with T-REX. J. Syst. Softw.
85(8), 1709–1728 (2012)
7. Panigati, E., Schreiber, F.A., Zaniolo, C.: Data streams and data stream manage-
ment systems and languages. In: Colace, F., De Santo, M., Moscato, V., Picariello,
A., Schreiber, F., Tanca, L. (eds.) Data Management in Pervasive Systems, pp.
93–111. Springer, Cham (2015)
8. Marelli, M., Fortunato, M., Camplani, R., Schreiber, F.A., Rota, G.: Perla: a lan-
guage and middleware architecture for data management and integration in per-
vasive information systems. IEEE Trans. Software Eng. 38, 478–496 (2012)
9. Colace, F., De Santo, M., Greco, L., Moscato, V., Picariello, A.: Probabilistic
approaches for sentiment analysis: latent dirichlet allocation for ontology building
and sentiment extraction. In: Pedrycz, W., Chen, S.M. (eds.) Sentiment Analysis
and Ontology Engineering, pp. 75–91. Springer, Cham (2016)
10. Bolchini, C., Quintarelli, E., Tanca, L.: Carve: context-aware automatic view def-
inition over relational databases. Inf. Syst. 38(1), 45–67 (2013)
11. Dong, X.L., Srivastava, D.: Big data integration. In: 2013 IEEE 29th International
Conference on Data Engineering (ICDE), pp. 1245–1248. IEEE (2013)
12. Amato, F., Castiglione, A., Moscato, V., Picariello, A., Sperlı́, G.: Multimedia
summarization using social media content. Multimed. Tools Appl. pp. 1–25 (2018)
13. Amato, F., Castiglione, A., De Santo, A., Moscato, V., Picariello, A., Persia,
F., Sperlı́, G.: Recognizing human behaviours in online social networks. Comput.
Secur. 74, 355–370 (2018)
14. Amato, F., Moscato, V., Picariello, A., Sperlı́, G.: Extreme events management
using multimedia social networks. Future Gener. Comput. Syst. 94, 444–452 (2019)
15. Bartolini, I., Patella, M.: Multimedia queries in digital libraries. In: Colace, F.,
De Santo, M., Moscato, V., Picariello, A., Schreiber, F., Tanca, L. (eds.) Data
Management in Pervasive Systems, pp. 311–325. Springer, Cham (2015)
16. Colace, F., De Santo, M., Greco, L., Moscato, V., Picariello, A.: A collabora-
tive user-centered framework for recommending items in online social networks.
Comput. Hum. Behav. 51, 694–704 (2015)
The 2nd International Workshop on
Internet of Everything and Machine
Learning Applications (IOEMLA-2019)
Social Credibility Incorporating
Semantic Analysis and Machine Learning:
A Survey of the State-of-the-Art and Future
Research Directions
Abstract. The wealth of Social Big Data (SBD) represents a unique opportu-
nity for organisations to obtain the excessive use of such data abundance to
increase their revenues. Hence, there is an imperative need to capture, load,
store, process, analyse, transform, interpret, and visualise such manifold social
datasets to develop meaningful insights that are specific to an application’s
domain. This paper lays the theoretical background by introducing the state-of-
the-art literature review of the research topic. This is associated with a critical
evaluation of the current approaches, and fortified with certain recommendations
indicated to bridge the research gap.
1 Introduction
The dramatic increase in the social data impact - a testimony to our growing digital
lifestyles - has taken on industries and activities ranging from marketing and adver-
tising to intelligence gathering and political influence. Interpreting such a Big Data
island is momentous to bring new perspectives and to improve business practices, yet
this revolution is still in its infancy. It is easy to assume that social data revolution is
about quantity, but social data is more than just volume. In fact, the extents of this
revolution are more wide-spreading: it is about building data infrastructures that are
needed to effectively digest the breeding of social data to infer the hoped-for added
value. This has motivated the research communities to dig deep, provide solutions and
to implement platforms into potential usage of these datasets to benefit several appli-
cations [1–7].
In modern enterprises, social networks are used as part of the infrastructure for a
number of emerging applications such as recommendation and reputation systems. In
such applications, trust is one of the most important factors in decision making.
Sherchan et al. [8] defined Trust as the measurement of confidence where a group of
individuals or communities behave predictably.
In the latter part of the 20th century, researchers in the field of Artificial Intelligence
(AI) have become active in computational modelling of statistical analysis means
alongside with defining ontologies that would deliver automated reasoning capabilities.
This section aims to review the existing approaches into two main categories:
(i) techniques incorporating semantic analysis for domain discovery; (ii) machine
learning statistical techniques for data classification topics distillation.
The previous sections present an integrated review of the existing methods and
approaches in the era of SBD incorporating trust, semantic analysis and machine
learning.
• Lack of advanced domain-based trustworthiness approaches,
• Lack of managing and extracting high-level domains from the textual content of
SBD,
• Lack of domain-based techniques for dual classification.
1
http://www.gartner.com/document/3383817?ref=solrAll&refval=175496307&qid=
34ddf525422cc71383ee22c858f2238a, Visited in 25/10/2016.
Social Credibility Incorporating Semantic Analysis and Machine Learning 891
5 Conclusion
Currently, several research efforts are undertaken to handle and manage the large scale
of SBD in the quest for added value. A thread of these efforts attempts to provide
technical solutions to cope with the volume and speed of social data generation. This is
through facilitating the process of data capturing, acquisition and storing. Another host
of scholars aim to develop data analytics solutions to enhance the quality of the collected
social data, to understand the users’ topics of interest, to classify and categorise users
into segments, and to extract better insights, thus improving the decision-making pro-
cess. Despite the significance of these endeavours and initiatives to change the per-
ception of the extracted social data, these current frameworks for SBD analysis only
partially consider SBD features. Furthermore, existing efforts are conducted merely to
address one or few issues of SBD. A comprehensive framework is required to resolve
the issues of data quality, extract hidden knowledge, and infer the credibility of users
and their content of the OSNs through extracting the domains of interest at the user and
the post level. This consolidates the development of data classification techniques which
leads to better anticipating of users’ interest in their future published content.
In this paper, an extensive survey of the existing literature is conducted. The
discussion of several approaches to measuring the generic-based and domain-based
trustworthiness in SBD is carried out followed by addressing the core literature in the
areas of semantic analysis, machine learning and data classification within the context
of SBD. The carried out efforts within the cordon of these themes are critically
reviewed which points out to the considerable achievements have been made in SBD
analytics. However, the current approaches and techniques are still insufficient in terms
of (i) the lack of domain-based trustworthiness approaches; (ii) the lack of managing
and extracting high-level domains from the textual content of SBD; (iii) the lack of
domain-based approaches for dual classification of the textual content of SBD.
894 B. Abu-Salih et al.
References
1. Abu-Salih, B., Wongthongtham, P., Kit, C.Y.: Twitter mining for ontology-based domain
discovery incorporating machine learning. J. Knowl. Manag. 22(5), 949–981 (2018)
2. Nabipourshiri, R., Abu-Salih, B., Wongthongtham, P.: Tree-based classification to users’
trustworthiness in OSNs In: Proceedings of the 2018 10th International Conference on
Computer and Automation Engineering, pp. 190–194. ACM, Brisbane, Australia (2018)
3. Wongthongtham, P., Abu Salih, B.: Ontology and trust based data warehouse in new
generation of business intelligence: state-of-the-art, challenges, and opportunities. In: 2015
IEEE 13th International Conference on Industrial Informatics (INDIN). IEEE (2015)
4. Abu-Salih, B., et al.: Towards a methodology for social business intelligence in the era of big
social data incorporating trust and semantic analysis. In: Second International Conference on
Advanced Data and Information Engineering (DaEng-2015). Springer, Bali, Indonesia
(2015)
5. Abu-Salih, B., et al.: CredSaT: credibility ranking of users in big social data incorporating
semantic analysis and temporal factor. J. Inform. Sci. 0(0), 0165551518790424 (2018)
6. Abu-Salih, B.: Trustworthiness in Social Big Data Incorporating Semantic Analysis,
Machine Learning and Distributed Data Processing. Curtin University (2018)
7. Chan, K.Y., et al.: Affective design using machine learning: a survey and its prospect of
conjoining big data. Int. J. Comput. Integr. Manuf. 1–19 (2018)
8. Sherchan, W., Nepal, S., Paris, C.: A survey of trust in social networks. ACM Comput. Surv.
45(4), 47 (2013)
9. Passant, A., et al.: Enabling trust and privacy on the social web. In: W3C Workshop on the
Future of Social Networking (2009)
10. Podobnik, V., et al.: How to calculate trust between social network users? In: 2012 20th
International Conference on Software, Telecommunications and Computer Networks
(SoftCOM). IEEE (2012)
11. Agarwal, M., Bin, Z.: Detecting malicious activities using backward propagation of
trustworthiness over heterogeneous social graph. In: 2013 IEEE/WIC/ACM International
Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies
(IAT) (2013)
12. Massa, P., Bhattacharjee, B.: Using trust in recommender systems: an experimental analysis,
in trust management. In: Jensen, C., Poslad, S., Dimitrakos, T. (eds.) pp. 221–235. Springer,
Berlin, Heidelberg. (2004)
13. Gupta, P., et al.: WTF: the who to follow service at Twitter. In: Proceedings of the 22nd
International Conference on World Wide Web, pp. 505–514. International World Wide Web
Conferences Steering Committee, Rio de Janeiro, Brazil (2003)
14. Gallege, L.S., et al.: Towards trust-based recommender systems for online software services.
In: Proceedings of the 9th Annual Cyber and Information Security Research Conference,
pp. 61–64. ACM, Oak Ridge, Tennessee (2014)
15. Sun, G., et al.: A social trust path recommendation system in contextual online social
networks. In: Chen, L., et al. (eds.) Web Technologies and Applications, pp. 652–656.
Springer, Cham (2014)
16. Alahmadi, D.H., Zeng, X.J.: ISTS: implicit social trust and sentiment based approach to
recommender systems. Expert Syst. Appl. 42(22), 8840–8849 (2015)
17. AlRubaian, M., et al.: A multistage credibility analysis model for microblogs. In:
Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social
Networks Analysis and Mining 2015. ACM (2015)
Social Credibility Incorporating Semantic Analysis and Machine Learning 895
18. Zhang, B., et al.: A trust-based sentiment delivering calculation method in microblog. Int.
J. Serv. Technol. Manag. 21(4–6), 185–198 (2015)
19. Bae, Y., Lee, H.: Sentiment analysis of twitter audiences: measuring the positive or negative
influence of popular Twitterers. J. Am. Soc. Inform. Sci. Technol. 63(12), 2521–2535 (2012)
20. Kawabe, T., et al.: Tweet credibility analysis evaluation by improving sentiment dictionary.
In: 2015 IEEE Congress on Evolutionary Computation (CEC). IEEE (2015)
21. Momeni, E., Cardie, C., Diakopoulos, N.: A survey on assessment and ranking methodolo-
gies for user-generated content on the web. ACM Comput. Surv. (CSUR) 48(3), 41 (2016)
22. Amalanathan, A., Anouncia, S.M.: A review on user influence ranking factors in social
networks. Int. J. Web Based Commun. 12(1), 74–83 (2016)
23. Ruan, Y., Durresi, A.: A survey of trust management systems for online social communities–
trust modeling, trust inference and attacks. Knowl.-Based Syst. 106, 150–163 (2016)
24. Berners-Lee, T., Hendler, J.: Publishing on the semantic web. Nature 410(6832),
1023 (2001)
25. Gruber, T.R.: Toward principles for the design of ontologies used for knowledge sharing?
Int. J. Hum Comput Stud. 43(5), 907–928 (1995)
26. De Nart, D., et al.: A content-based approach to social network analysis: a case study on
research communities, in digital libraries on the move. In: Calvanese, D., De Nart, D., Tasso,
C. (eds.) 11th Italian Research Conference on Digital Libraries, IRCDL 2015, Bolzano, Italy,
29–30 January 2015, pp. 142–154. Springer, Cham, 2016, Revised Selected Papers
27. Chianese, A., Marulli, F., Piccialli, F.: Cultural heritage and social pulse: a semantic
approach for CH sensitivity discovery in social media data. In: 2016 IEEE Tenth
International Conference on Semantic Computing (ICSC) (2016)
28. Michelson, M., Macskassy, S.A.: Discovering users; topics of interest on twitter: a first look.
In: Proceedings of the Fourth Workshop on Analytics for Noisy Unstructured Text Data.
ACM (2010)
29. Schonhofen, P.: Identifying document topics using the wikipedia category network. In:
Proceedings of the 2006 IEEE/WIC/ACM International Conference on Web Intelligence,
pp. 456–462. IEEE Computer Society (2006)
30. Hassan, M.M., Karray, F., Kamel, M.S.: Automatic document topic identification using
wikipedia hierarchical ontology. In: 2012 11th International Conference on Information
Science, Signal Processing and their Applications (ISSPA) (2012)
31. Anthes, G.: Topic models Vs unstructured data. Commun. ACM 53(12), 16–18 (2010)
32. Wang, C., et al. Markov topic models. In: Artificial Intelligence and Statistics (2009)
33. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3(4–5),
993–1022 (2003)
34. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3(Jan),
993–1022 (2003)
35. Karami, A., et al.: Fuzzy approach topic discovery in health and medical corpora. Int.
J. Fuzzy Syst. 1–12 (2017)
36. Hofmann, T.: Probabilistic latent semantic indexing. In: Proceedings of the 22nd Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval. ACM (1999)
37. Chen, Y., et al.: Topic modeling for evaluating students; reflective writing: a case study of
pre-service teachers’ journals. In: Proceedings of the Sixth International Conference on
Learning Analytics & Knowledge. ACM (2016)
38. Nichols, L.G.: A topic model approach to measuring interdisciplinarity at the national
science foundation. Scientometrics 100(3), 741–754 (2014)
39. Weng, J., et al.: Twitterrank: finding topic-sensitive influential Twitterers. In: Proceedings of
the Third ACM International Conference on Web Search and Data Mining. ACM (2010)
896 B. Abu-Salih et al.
40. Asharaf, S., Alessandro, Z.: Generating and visualizing topic hierarchies from microblogs:
an iterative Latent Dirichlet allocation approach. In: 2015 International Conference on
Advances in Computing, Communications and Informatics (ICACCI). IEEE (2015)
41. Quercia, D., Askham, H., Crowcroft, J.: TweetLDA: supervised topic classification and link
prediction in Twitter. In: The 4th Annual ACM Web Science Conference, pp. 247–250.
ACM, Evanston, Illinois (2012)
42. Onan, A., Korukoglu, S., Bulut, H.: LDA-based topic modelling in text sentiment
classification: an empirical analysis. Int. J. Comput. Linguistics Appl. 7(1), 101–119 (2016)
43. Cha, M., et al.: Measuring user influence in Twitter: the million follower fallacy. ICWSM
10, 10–17 (2010)
44. Silva, A., et al.: ProfileRank: finding relevant content and influential users based on
information diffusion. In: Proceedings of the 7th Workshop on Social Network Mining and
Analysis. ACM (2013)
45. Jiang, W., Wang, G., Wu, J.: Generating trusted graphs for trust evaluation in online social
networks. Future Gener. Comput. Syst. 31, 48–58 (2014)
46. Liu, B., Zhang, L.: A survey of opinion mining and sentiment analysis. In: Aggarwal, C.,
Zhai, C. (eds.) Mining Text Data, pp. 415–463. Springer, Cham (2012)
47. Balog, K., et al.: Expertise retrieval. Found. Trends Inf. Retr. 6(2–3), 127–256 (2012)
48. Yin, H.Z., et al.: Dynamic user modeling in social media systems. ACM Trans. Inf. Syst.
33(3), 10 (2015)
49. Abbasi, M.-A., Liu, H.: Measuring user credibility in social media, in social computing,
behavioral-cultural modeling and prediction. In: Greenberg, A., Kennedy, W., Bos, N. (eds.)
pp. 441–448. Springer, Berlin, Heidelberg (2013)
50. Yardi, S., et al.: Detecting spam in a Twitter network (2009)
51. Manyika, J., et al.: Big data: the next frontier for innovation, competition, and productivity.
McKinsey Global Institute (2011)
52. Chen, M., et al.: Open issues and outlook. In: Big Data, pp. 81–89. Springer International
Publishing (2014)
53. Lavbič, D., et al.: Traversal and relations discovery among business entities and people using
semantic web technologies and trust management. In: Databases and Information Systems
VII: Selected Papers from the Tenth International Baltic Conference, DB & IS 2012. IOS
Press (2013)
54. Herzig, J., Mass, Y., Roitman, H.: An author-reader influence model for detecting topic-
based influencers in social media. In: Proceedings of the 25th ACM Conference on
Hypertext and Social Media. ACM (2014)
55. Shen, W., Wang, J., Han, J.: Entity linking with a knowledge base: Issues, techniques and
solutions. IEEE Trans. Knowl. Data Eng. 27(2), 443–460 (2015)
56. Free Social Media Analytics Tools (2016). http://simplymeasured.com/free-social-media-
tools/
57. Li, C., et al.: Topic Modeling for short texts with auxiliary word embeddings. In:
Proceedings of the 39th International ACM SIGIR Conference on Research and
Development in Information Retrieval. ACM (2016)
Source Codes Classification Using a Modified
Instruction Count Pass
1 Introduction
The vulnerability is defined as a weakness or flaw in systems that make the systems
easy to be attacked or reduce its security level [1]. There are several types of vulner-
ability, some of them are related to memory such as; memory leak, null pointers and
buffer overflow. We concentrate on this project in the first two types. The problem of
memory leak is the allocation problem of some memory areas that didn’t release when
they no longer needed. For example, in OOP, such an object may be stored in a
memory and cannot be accessed by the running code. This memory area is allocated for
something unusable. A null pointer is defined as a pointer which points to the invalid
object. In other words, attempting to read from or write to a memory area which not
mapped or no longer used. This may lead to read rubbish values or null values. In both
cases, a crash may occur.
© Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 897–906, 2019.
https://doi.org/10.1007/978-3-030-15035-8_88
898 O. Darwish et al.
Managing this problem can be done merely by manually auditing the source codes.
This approach may be valid if the code consists of a few lines or you have just a few
small codes. Moreover, this process needs good experience in programming and
software design. The security background is significant as well. This process is time-
consuming, and its results depend on the auditor and his/her skills. Due to the above,
other automatic techniques were introduced and now used to detect or predict the
existence of vulnerability. These techniques can be either static or dynamic. In the latter
one, an auditor will run the program and examine its behaviour through the runtime. In
[2], they provide a checking code technique with GCC compiler to detect those
potentially unsafe null pointers and memory leak. Static analysis is another technique
to detect the vulnerability.
In comparison with manual auditing, it can be efficient and save much time. In
general, the idea behind this technique is to depend on the predefined knowledge base
that contains keywords or functions which may cause the bug in the software. How-
ever, in practice, this technique has many limitations and disadvantages. The high false
positive is one of the weaknesses in tools which used static analysis. The generated
reports contain many unsafe programs, but in reality, this is not true. A flawfinder and
LLVM clang static analyzers are the example of those tools [3, 4].
2 LLVM Environment
This pass is one of many passes developed by LLVM team. It is an analysis pass and
doesn’t change the source code but gives some statistics about the code. These values
are so important and give the programmer an essential guide about his/her code exe-
cution. In this section, the output of this pass will be discussed in details.
different basic block based on condition(s). This happens when, for example, a con-
ditional statement used in the code.
In the same fashion, a ‘switch’ instruction can be used. The difference between ‘br’
and ‘switch’ is that switch can use many ‘br’ as it transfers the execution from one
place to another. Similarly, the ‘invoke’ instruction causes control to transfer to the
specified function, with the possibility of control flow transfer to either the
‘normal’ label or to ‘exception’ label. The normal label will be used if the caller uses
the ‘ret’ instruction. Without a successor, ‘resume’ is termination instruction.
If a source code contains a statement that will not be executed anyway, this means
these statements are unreachable. ‘Unreachable’ instructions can be counted via this
pass. This is a good indication that these statements should be removed or managed in
such a way.
4 Related Work
Mozilla Firefox showed that it is possible in around 72% of the time. This is based on
the classifier built using decision tree. Hata, Mizuno et al. [9], provide an approach
using text analysis to predict flaws in software. They applied the approach on five open
source eclipse projects and got just 40% of precision and 80% of recall. Using text
mining, Aversano et al. [10] built a predictor to check the newly added code in such
software contains bug or not. The KNN technique was used and applied over two open
java source. The built tool got 59% precision and 69% of recall. The above is a sample
of the related work in bug prediction in software. All of them use the high-level source
code analysis to build a tool for vulnerability detection. [11] performed an early study
to predict vulnerability using text mining on the source code. Our work has the same
target, but we try to do the same thing using low-level codes.
5 Methodology
For the research purposes, 1000 source codes written in C and C++ language are
collected from different sources. We use the benchmark of NIST [12], and around 700
codes from math library and planet source archives websites [13]. These codes are
compiled using Clang built in Ubuntu 14.0. There are many source codes contains
syntax errors. We try to fix them, but it is time-consuming and needs too much effort as
some of them include more than one error. In our dataset, we use 250 benchmark codes
from NIST and other 250 codes from the other sources. To make sure that the latter 250
are free of bugs, we have used the flawfinder tool, and we found that most of them are
clean. The buggy codes are replaced by others where the flawfinder decide those as
clean codes. The second step in this research is to generate the bitcodes of the dataset.
We use LLVM commands for this issue. After that, the “instcount” pass was modified
to get a specific output format, then applied over the dataset. Every source code will get
another txt file contains its analysis results. An around 250 LOC matLab analyzer was
built to collect the files and parse the content and unify them in one csv file to be ready
for data mining WEKA tool. Figure 1 below shows the processes of this research.
The instcount model was trained on 66% of the 500 records and tested on the other
ones. These numbers are default values in WEKA. A decision tree generated from
WEKA tool was implemented. It contains 12 attributes from originally 40 attributes
that WEKA found them as good features to classify the 500 source codes into three
classes; clean, the vulnerability in the pointer, and vulnerability of memory leak. It is
clear that the classification process depends mainly on the values of basic blocks and
Phi. Also, the unreachable instructions have a major role in building the decision tree.
The confusion matrix in Fig. 2 (Appendix) shows that the classifier accuracy is around
88%. This means that this classifier can be applied in real applications. The classifi-
cation process based on the low level gives a strong indication that we can predict the
existence of vulnerability based on the relation between the instructions counts and
values. In comparison with static analysis tools, we found that our classifier super
performed the LLVM Clang static analyzer. We use all the benchmark codes from
NIST to assess our classifier, flawfinder, and clang analyzer. Table 1 shows that LLVM
clang analyzer needs more effort to improve its accuracy and efficiency in vulnerability
detection. We also apply the WEKA decision tree technique to classify the vulnerable
software itself into two classes we are working on; null pointer and memory. The
classification process showed that we could build an efficient classifier to differentiate
between vulnerability types. Figure 3 (Appendix) indicates that the constructed clas-
sifier gets around 94% of accuracy. After building the decision tree and generate the
rules, the InstCount pass in LLVM was modified, and the rules were added. The output
of the new version contains the decision of the classifier about the vulnerability exists
in the input source code. Figure 4-a, b, c (Appendix) shows the output of sample files
of source codes.
In this research, we spent too much time in dataset collecting. Many difficulties are
encountered because many of the downloaded files have a syntax error(s). Some file
can be compiled in the windows environment, but not in Linux, as they contain some
pure windows header files and they need the effort to convert them into Linux like
environment. In order to get reliable results, the dataset should be as much as possible.
The collecting process is not easy and needs effort and time.
In this project, we exploit the results of LLVM pass to predict the existence of vul-
nerable software. We depend on the low-level instructions. We use WEKA tool to
build a decision tree and generate the rules. Our classifier has 88% of accuracy. Based
on the results, the classifier super performed the LLVM Clang Static Analyzer in
prediction the buggy source codes. However, both of them are using a different
technique; we argue that our classifier can be considered as enhancement of the Clang
analyzer and we can use both of them together for more efficient results.
For the future plan, we will study these results in depth to find the relation among
the instruction numbers that gives the pattern of buggy codes. Also, we are going to
enrich the dataset for both vulnerable and clean source codes. Moreover, we will use
advanced data mining techniques to find more strong rules based on the instructions.
904 O. Darwish et al.
Appendix
(a)
(b)
(c)
Fig. 4. Sample of the output of new instcount pass.
906 O. Darwish et al.
References
1. Hughes, J., Cybenko, G.: Three tenets for secure cyber-physical system design and
assessment. In: Cyber Sensing 2014, vol. 9097, p. 90970A. International Society for Optics
and Photonics (2014)
2. Eigler, F.C.: Mudflap: Pointer use checking for C/C++ (2003)
3. Flawfinder. http://www.dwheeler.com/flawfinder/
4. Clang Static Analyzer. http://clang-analyzer.llvm.org/
5. The LLVM Compiler. Infrastructure. http://llvm.org/
6. Viega, J., Bloch, J.T., Kohno, Y., McGraw, G.: Its4: a static vulnerability scanner for C and
C++ code. In: 16th Annual Conference on Computer Security Applications, ACSAC 2000,
pp. 257–267. IEEE (2000)
7. Shin, Y., Williams, L.: Is complexity really the enemy of software security? In: Proceedings
of the 4th ACM Workshop on Quality of Protection, pp. 47–50. ACM (2008)
8. Chowdhury, I., Zulkernine, M.: Using complexity, coupling, and cohesion metrics as early
indicators of vulnerabilities. J. Syst. Architect. 57(3), 294–313 (2011)
9. Mizuno, O., Ikami, S., Nakaichi, S., Kikuno, T.: Spam filter based approach for finding fault-
prone software modules. In: Proceedings of the Fourth International Workshop on Mining
Software Repositories, p. 4. IEEE Computer Society (2007)
10. Aversano, L., Cerulo, L., Del Grosso, C.: Learning from bug-introducing changes to prevent
fault prone code. In: Ninth International Workshop on Principles of Software Evolution: in
Conjunction with the 6th ESEC/FSE Joint Meeting, pp. 19–26. ACM (2007)
11. Scandariato, R., Walden, J., Hovsepyan, A., Joosen, W.: Predicting vulnerable software
components via text mining. IEEE Trans. Softw. Eng. 40(10), 993–1006 (2014)
12. National Institute of Standards and Technology. http://www.nist.gov
13. Mathematics Source Library C & ASM. http://www.mymathlib.com
Predictive Analytics and Deep Learning
Techniques in Electronic Medical Records:
Recent Advancements and Future Direction
Abstract. The demands on medical services are increasing rapidly in the global
context. Therefore, handling beds availability, identifying and managing the
length of stay (LOS) is creating persistent needs for the physicians, nurses,
clinicians, hospital management, and caregivers in the public hospital admis-
sions and the private hospital admissions. Health analytics provides unprece-
dented ways to predict trends, patients’ future outcomes, knowledge discovery,
and improving the decision making in the clinical settings. This paper reviews
the state-of-the-art machine learning, deep learning techniques and the related
work in relation to the length of stay common hospital admissions. Research
trends and future direction for the forecasting LOS in medical admissions are
discussed in this paper.
1 Introduction
departments (e.g. Acute myocardial infarction in the ED). Machine learning techniques
are expediting the processes for the artificial intelligence (AI) in the clinical field and
helping the researchers in this domain to construct predictive medical modelling and
health analytical tools that can support clinical decision making.
Machine learning and recently Deep Learning models can work effectively with
EHRs in an automatic and collective process [3]. As a result, this will help the
researchers and the data scientists to collect massive records of data, including patients’
health records, behavioral data, temporal data and time series data; therefore, better
intelligent diagnostic predictions from the data can be achieved. Predicting the number
of days which patients are required to stay at hospitals can be beneficial for hospitals,
patients and caregivers. Consequently, this will lead to preventing the late discharge or
early discharge for particular medical problems that require a prolonged LOS. In
prediction learning modellings, whether supervised learning or unsupervised models
can best predict the particular medical case, or even hybrid approach it is necessary to
deal with the complex medical case that requires appropriate data representation prior
to the prediction process.
The motivation of this work relies on the fact that highlighting and discussing the
most recent advancement in predictive analytics algorithms may lead to a better
understanding of current techniques of the AI in LOS EHR predictive tasks. This
comprises studying these techniques in the literature review, research challenges and
the future research direction for forecasting LOS predictions problems. Therefore, this
paper aims to summarize the state-of-the-art machine learning and deep learning
techniques related works for predicting LOS in common hospital admissions. These
works are group into five main categories. Acute Hospital Admission, Intensive Care
Unit (ICU), Emergency Department ED, Pediatric Emergency Department (PED),
General Hospital Admission, and Cardiac Surgery.
Many research projects [4, 5] have endeavored to address LOS prediction at the
emergency department, ICU and the different types of hospital admission considering
LOS as a prediction problem using regression analysis models, while other research
works [6–8] have taken into account the classification models in the LOS prediction
problems, meanwhile they are fewer types of research exploiting clustering-based
models in LOS. These models were examined in various clinical settings for fore-
casting of length stay tasks and to find the best predictive technique in Acute Medicine,
Intensive Care Unit (ICU), Emergency Department ED, Pediatric Emergency Depart-
ment (PED), General Hospital Admission, and Cardiac Surgery. These are the common
predictive tasks in forecasting LOS in hospitals admission. Each predictive task has
different affecting factors which may influence the average length of stay each one
(short LOS, medium LOS, and long LOS or “prolonged LOS”).
Predictive Analytics and Deep Learning Techniques 909
Table 2. Represents a comparison between the common supervised learning models in LOS
Hospital Admissions common predictions problems.
Classification Example Model advantages Model disadvantages
model
Support [6, 17] – Fast to predict outcomes – Not efficient in large
Vector – Efficient in non-linear dataset (require a lot of time
Machine separation problems for training data)
(SVM) – Computation speed and – Not efficient when target
memory classes are overlapping
K-Nearest [19] – Simple algorithm and – Features selection problem
Neighbors easy to implement – Takes a lot of memory and
(kNN) – Learning is simple CPU to calculate nearest
(almost no learning at all) neighbors in a large dataset
Naive Bayes [20] – Fast to predict – Not suitable if categorical
– Useful in multi-class variable has Zero probability
prediction – Not efficient if predators are
– Efficient in categorical not completely independent
input variables
Random [8, 17, 18] – Useful in non-linear – Computational cost
Forest prediction problems – Takes time to build and
– Useful in feature create
engineering (can find
important features in trained
dataset)
techniques can engineer features which are suitable for many clinical and medical tasks
or customize specific features for specific tasks [21]. In general, deep learning based
models require large data to work efficiently. In other words, they need enough data to
train deep learning models in various prediction problems.
Recurrent Neural Network (RNN) is a common deep learning technique. RNN
(Fig. 1) is a type of neural networks which is an appropriate choice when clinical data
is sequentially ordered (time-series or natural language), especially that RNNs are
designed to deal with such long-range temporal dependencies [22]. Recently,
researcher started to exploit the immense benefits of deep learning based techniques in
EMR context. For example, Che et al. [23] applied RNN in time-series classification on
multiple clinical prediction tasks (including the length of stay) extracted from three
openly available databases (UCI, MIMIC III, and PhysioNet). Similarly, Harutyunyan
et al. [24] applied RNN using public database (MIMIC III) to predict the length of stay
in the problem of the multiple time series event in clinical settings.
The LSTM neural network is a type of the RNN which can learn order dependences
in sequential prediction problems. The LSTM consists of three gates (Eqs. 1, 2, and 3);
the first gate is the forget state (ft), the second gate is the input gate (It) and the output
gate (Ot), where each gate has different sets of weights. The following equations
represent the three gates and three states:
Numerous research studied the LSTM performance in the clinical domain. For
instance, Fang et al. [1] compared the performance of class models (Logistic Regres-
sion (LR) and Feed-Forward (MLP)) with LSTM, and Multitask LSTM models in LOS
prediction task. The results indicated that LSTM and Multitask LSTM outperformed
traditional learning (shallow learning). Comparably, Rajkomar el at [25] applied deep
learning LSTM to predict the length of stay of hospitalized patients in the first 24 h of
admission.
912 B. Alsinglawi and O. Mubin
3 Future Direction
With the recent advancement in deep learning models in the medical domain, however,
deep learning models are still in the early development stage in EMR context with the
focus on the length of stay in the common hospital admission prediction tasks. Since
there are little few attempts to take advantage of the deep learning techniques in this
field, the future directions and trends can be pointed to the following research interest:
– Maximize the benefits of the current models such as DNN, RNN, LSTM and
implements Unsupervised (e.g. Deep Belief Network, Autoencoder, etc.) and Rein-
forcement deep learning models to ease and build a more efficient model to achieve
promising results in various hospital admission prediction tasks including hospital
length of stay, discharge, the risk of re-admission, and future patients’ outcomes.
– Deep learning based techniques have tremendous benefits for Big Data analytics
such the use of deep neural network layers of hidden variables, and the automatic
features learning (pre-processing, feature, extraction and feature selection.). In the
domain of deep learning and big data health analytics, there are common techniques
which were recently adopted in hospital admissions and prediction tasks such as the
LSTM and RNN. Recently, researchers started to exploit the benefits of deep
learning unsupervised techniques in big data analytics such as (Auto-encoders,
Boltzmann Machines (RBMs), and Deep Belief Networks (DBN)). Therefore, the
adoption of big data analytics remains unexplored research area in the domain of
EHR for the hospital length of stay prediction problems.
Predictive Analytics and Deep Learning Techniques 913
4 Conclusion
Previous research tells us identifying the right predictive learning models that can suit
specific medical cases in forecasting length of stay for is rendered as a significant task
in hospital admissions. Certainly, this is considered as a research challenge especially
in deep learning models and big data analytics which require further investigations to
provide promising outcomes to forecast hospital length of stay for the target medical
prediction task.
References
1. Tang, F., Xiao, C., Wang, F., Zhou, J.: Predictive modeling in urgent care: a comparative
study of machine learning approaches. JAMIA Open (2018)
2. What is an electronic health record (HER) (2018). https://www.healthit.gov/faq/what-
electronic-health-record-ehr
3. Gunasekar, S., Ho, J.C., Ghosh, J., Kreml, S., Kho, A.N., Denny, J.C., Malin, B.A., Sun, J.:
Phenotyping using structured collective matrix factorization of multisource ehr data. arXiv
preprint arXiv:1609.04466 (2016)
4. Liu, V., Kipnis, P., Gould, M.K., Escobar, G.J.: Length of stay predictions: improvements
through the use of automated laboratory and comorbidity variables. Medical care, pp. 739–
744 (2010)
5. Hosseininejad, S.M., Aminiahidashti, H., Pashaei, S.M., Khatir, I.G., Montazer, S.H.,
Bozorgi, F., Mahmoudi, F.: Determinants of prolonged length of stay in the emergency
department; a cross-sectional study. Emergency 5(1) (2017)
6. Hachesu, P.R., Ahmadi, M., Alizadeh, S., Sadoughi, F.: Use of data mining techniques to
determine and predict length of stay of cardiac patients. Healthc. Inform. Res. 19(2),
121–129 (2013)
7. Maharlou, H., Niakan Kalhori, S.R., Shahbazi, S., Ravangard, R.: Predicting length of stay
in intensive care units after cardiac surgery: comparison of artificial neural networks and
adaptive neuro-fuzzy system. Healthc. Inform. Res. 24(2), 109–117 (2018)
8. Elbattah, M., Molloy, O.: Using machine learning to predict length of stay and discharge
destination for hip-fracture patients. In: Proceedings of SAI Intelligent Systems Conference,
pp. 207–217. Springer (2016)
9. Azari, A., Janeja, V.P., Levin, S.: Imbalanced learning to predict long stay emergency
department patients. In: 2015 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM), pp. 807–814. IEEE (2015)
10. Vermeulen, M.J., Guttmann, A., Stukel, T.A., Kachra, A., Sivilotti, M.L., Rowe, B.H.,
Dreyer, J., Bell, R., Schull, M.: Are reductions in emergency department length of stay
associated with improvements in quality of care? a difference-in-differences analysis. BMJ
Qual. Saf. 25(7), 489–498 (2016)
914 B. Alsinglawi and O. Mubin
11. Barak-Corren, Y., Fine, A.M., Reis, B.Y.: Early prediction model of patient hospitalization
from the pediatric emergency department. Pediatrics 139(5), e20162785 (2017)
12. Carter, E.M., Potts, H.W.: Predicting length of stay from an electronic patient record system:
a primary total knee replacement example. BMC Med. Inform. Decis. Making 14(1), 26
(2014)
13. Zaccardi, F., Webb, D.R., Davies, M.J., Dhalwani, N.N., Gray, L.J., Chatterjee, S., Housley,
G., Shaw, D., Hatton, J.W., Khunti, K.: Predicting hospital stay, mortality and readmission
in people admitted for hypoglycaemia: prognostic models derivation and validation.
Diabetologia 60(6), 1007–1015 (2017)
14. Juma, S., Taabazuing, M.M., Montero-Odasso, M.: Clinical frailty scale in an acute
medicine unit: a simple tool that predicts length of stay. Can. Geriatr. J. 19(2), 34 (2016)
15. Magalhães, T., Lopes, S., Gomes, J., Seixo, F.: The predictive factors on extended hospital
length of stay in patients with ami: laboratory and administrative data. J. Med. Syst. 40(1), 2
(2016)
16. Cereda, E., Klersy, C., Pedrolli, C., Cameletti, B., Bonardi, C., Quarleri, L., Cappello, S.,
Bonoldi, A., Bonadeo, E., Caccialanza, R.: The geriatric nutritional risk index predicts
hospital length of stay and in-hospital weight loss in elderly patients. Clin. Nutr. 34(1),
74–78 (2015)
17. Combes, C., Kadri, F., Chaabane, S.: Predicting hospital length of stay using regression
models: application to emergency department. In: 10ème Conférence Francophone de
Modélisation, Optimisation et Simulation-MOSIM’14 (2014)
18. Chuang, M.T., Hu, Y.H., Lo, C.L.: Predicting the prolonged length of stay of general
surgery patients: a supervised learning approach. Int. Trans. Oper. Res. 25(1), 75–90 (2018)
19. Salah, H.: Predicting Inpatient Length of Stay in Western New York Health Service Area
Using Machine Learning Algorithms. Ph.D. thesis, State University of New York at
Binghamton (2017)
20. Muhlestein, W.E., Akagi, D.S., Davies, J.M., Chambless, L.B.: Predicting inpatient length of
stay after brain tumor surgery: developing machine learning ensembles to improve predictive
performance. Neurosurgery (2018)
21. Ching, T., Himmelstein, D.S., Beaulieu-Jones, B.K., Kalinin, A.A., Do, B.T., Way, G.P.,
Ferrero, E., Agapow, P.M., Zietz, M., Hoffman, M.M., et al.: Opportunities and obstacles for
deep learning in biology and medicine. J. R. Soc. Interface 15(141) (2018)
22. Shickel, B., Tighe, P., Bihorac, A., Rashidi, P.: Deep EHR: a survey of recent advances in
deep learning techniques for electronic health record (EHR) analysis. arXiv preprint arXiv:
1706.03446 (2017)
23. He, Z., Purushotham, S., Cho, K., Sontag, D., Liu, Y.: Recurrent neural networks for
multivariate time series with missing values. Sci. Rep. 8(1), 6085 (2018)
24. Harutyunyan, H., Khachatrian, H., Kale, D.C., Galstyan, A.: Multitask learning and
benchmarking with clinical time series data. arXiv preprint arXiv:1703.07771 (2017)
25. Rajkomar, A., Oren, E., Chen, K., Dai, A.M., Hajaj, N., Hardt, M., Liu, P.J., Liu, X.,
Marcus, J., Sun, M., et al.: Scalable and accurate deep learning with electronic health
records. NPJ Digital Med. 1(1), 18 (2018)
26. Dimitrov, D.V.: Medical internet of things and big data in healthcare. Healthc. Inform. Res.
22(3), 156–163 (2016)
Big Data Analytics for Electricity Price
Forecast
1 Introduction
One of the main applications of smart grid is the ability to provide the flexible load
curve and shift the load from the peak time. To be able to achieve this goal, dynamic
pricing would be a great help. Accurate price forecast can help the customers to make a
decision for their electricity consumption and adjust their electricity consumption
pattern based on accurate forecast. This quantitative and accurate decision making not
only helps the customers to optimize their electricity consumptions but also helps the
electricity market participants and the operators to run the market more efficiently. The
features which are going to influence the network are very diverse and having an
algorithm which can extract the important features are crucial to determine the price of
the electricity. The good news is that large installations of the smart meters provide
very comprehensive historical data which can be used for electricity price forecast in
short and long term [1–3].
During the current research the electricity price forecast researches was reviewed
and the various applied methods are analyzed. Summary of the sum of the analyzed
researches presented in the following paragraph. Many of the publications related to the
electricity price forecast applied time series analysis and the machine learning methods
for electricity price forecast. A hybrid model for predicting the electricity price is
implemented by Varshney to predict day ahead electricity market based on the
extracted featured such as temperature and demand information [4]. A probabilistic
methodology combined with the wavelet neural networks is utilized to forecast the
hourly price with considering the uncertainty [5]. A markov chain model was used to
calculate the electricity pricing based on welfare with considering a tradeoff between
the customers and the utility companies [6]. The hourly electricity price forecast is
improved by applying multilayer neural network. The applied neural network in [7]
achieved the promising result and improved the accuracy of the forecast.
Time series analysis also used extensively for the purpose of forecasting the
electricity price. One of the most famous methods used for times series analysis is
called auto-regressive integrated moving average (ARIMA) [8]. Feature engineering is
one of the key requirements to determine the electricity price as the accurate forecasting
the electricity price demands feeding the relevant historical data with most correlated
features to the electricity price. The research referenced in [9–11] applied multi-
variable mutual information, interaction gain as well as C4.5 algorithms for feature
extractions. Traditional studies applied the decision trees as well as artificial neural
networks and according to [12] and [13] the traditional methods are not able to leverage
the big data values for the price forecast very efficiently.
3 Numerical Study
power import and export to specific geographical area as well as undeniable role of
renewable energy in the network. There are also other major economical factors which
could potentially impact the electricity price such as gross domestic product (Fig. 2).
Fig. 1. The highly correlated points are shown by yellow points and least correlated features are
determined by the dark blue points.
The Fig. 3 demonstrates the correlation of the electricity price with the variation in
natural gas price in California. As it is shown by the feature selection algorithm as well,
the natural gas price has a positive correlation with the electricity price.
Big Data Analytics for Electricity Price Forecast 919
Fig. 2. Demonstrates the correlation value and the applied color coding
After preprocessing, S-ARIMA was used to synthetically augment the data from
the 11 selected features from 2001–2014 in order to create projected values to compare
with the 2014–2017 test set. Upon performing linear regression on the augmented
dataset, we obtained MSE = 0.015, MAE = 0.096 and RMSE = 0.125, training
accuracy as 84.9% and test accuracy as 64% which suggested some degree of over-
fitting. Narrowing down to the 6 most seasonal features and training several other
models yielded drastically improved results. The Fig. 4 demonstrates the forecast price
for the PG&E utility in California.
920 A. Yousefi et al.
Fig. 3. Demonstrates the variation of the PG&E electricity price with the gas price variation in
California
Fig. 4. Demonstrates the predicted price versus actual price in the PG&E utility are for duration
of three years.
Big Data Analytics for Electricity Price Forecast 921
4 Conclusion
Using the time series data method described in this research, a dynamic model capable
of changing weights and adding or adjusting the influence of features over a particular
time bracket in the future, and providing several possible outcomes based on qualitative
parameters like the probability of beneficial or adversarial political or economic events
occurring at a given time is developed. Conceivably, this model would utilize online
learning from databases updated in real time providing constant feedback into the
pipeline and would be able to translate real world economic changes to a corresponding
change in the feature set and weights used in the model. This would allow the user to
consider different scenarios by changing the contributing factors of the prediction.
References
1. Agrawal, R.K., Muchahary, F., Tripathi, M.M.: Long term load forecasting with hourly
predictions based on long-short-term-memory networks. In: 2018 IEEE Texas Power and
Energy Conference (TPEC), pp. 1–6 (2018)
2. Helseth, A., Haugen, M., Jaehnert, S., Mo, B., Farahmand, H., Naversen, C.: Multi-market
price forecasting in hydro-thermal power systems. In: 15th International Conference on the
European Energy Market (EEM), pp. 1–5 (2018)
3. Jiang, L. Hu, G.: Day-ahead price forecasting for electricity market using long-short term
memory recurrent neural network. In: 15th International Conference on Control, Automa-
tion, Robotics and Vision (ICARCV), pp. 949–954 (2018)
4. Varshney, H., Sharma, A., Kumar, R.: A hybrid approach to price forecasting incorporating
exogenous variables for a day ahead electricity Market. In: 2016 IEEE 1st International
Conference on Power Electronics, Intelligent Control and Energy Systems (ICPEICES),
pp. 1–6 (2016)
5. Rafiei, M., Niknam, T., Khooban, M.: Probabilistic forecasting of hourly electricity price by
generalization of ELM for usage in improved wavelet neural network. IEEE Trans. Industr.
Inf. 13(1), 71–79 (2017)
6. Kobayashi, K.: Optimal real-time pricing of electricity based on switched markov chain
models and welfare maximization. In: IEEE 4th International Conference on Cyber-Physical
Systems, Networks, and Applications (CPSNA), pp. 35–39 (2016)
7. Mosbah, H., El-hawary, M.: Hourly electricity price forecasting for the next month using
multilayer neural network. Can. J. Electr. Comput. Eng. 39(4), 283–291 (2016)
8. Ozozen, A., Kayakutlu, G., Ketterer, M., Kayalica, O.: A combined seasonal ARIMA and
ANN model for improved results in electricity spot price forecasting: Case study in Turkey.
In: 2016 Portland International Conference on Management of Engineering and Technology
(PICMET), pp. 2681–2690 (2016)
9. Zhi-Wei, Q.: Mutivariable mutual information based feature selection for electricity price
forecasting. In: 2012 International Conference on Machine Learning and Cybernetics, vol. 1,
pp. 168–173 (2012)
10. Abedinia, O., Amjady, N., Zareipour, H.: A new feature selection technique for load and
price forecast of electrical power systems. IEEE Trans. Power Syst. 32(1), 62–74 (2017)
11. Hehui, Q., Zhiwei, Q.: Feature selection using C4.5 algorithm for electricity price prediction.
In: 2014 International Conference on Machine Learning and Cybernetics, vol. 1, pp. 175–
180 (2014)
922 A. Yousefi et al.
12. Wang, K., et al.: A survey on energy internet: architecture, approach, and emerging
technologies. IEEE Syst. J. 12(3), 2403–2416 (2018)
13. Jiang, H., Wang, K., Wang, Y., Gao, M., Zhang, Y.: Energy big data: a survey. IEEE Access
4, 3844–3861 (2016)
14. California Government Open Data Portal (2017). https://data.ca.gov/
15. Python Data Analysis Library (2017)
16. Brownlee, J.: Autoregression Models for Time Series Forecasting With Python – Machine
Learning Mastery (2018)
Queue Formation Augmented with Particle
Swarm Optimisation to Improve Waiting Time
in Airport Security Screening
Abstract. Airport security screening processes are essential to ensure the safety
of both passengers and the aviation industry. Security at airports has improved
noticeably in recent years through the utilisation of state-of-the-art technologies
and highly trained security officers. However, maintaining a high level of
security can be costly to operate and implement. It may also lead to delays for
passengers and airlines. This paper proposes a novel queue formation method
based on a queueing theory model augmented with a particle swarm optimisa-
tion method known as QQT-PSO to improve the average waiting time in airport
security areas. Extensive experiments were conducted using real-world datasets
collected from Sydney airport. Compared to the existing system, our method
significantly reduces the average waiting time and operating cost by 11.89%
compared to the one-queue formation.
The aviation industry continues to grow every year and has become an essential mode
of passenger transportation. In airports all around the world, there has been a contin-
uous increase in the number of passengers. In 2012, for example, 37 million passengers
travelled through Sydney airport, and by 2033 this number is estimated to increase to
74 million passengers. It is also estimated that by 2033, the number of passengers
travelling by air worldwide will increase to 6 billion [1]. In this context, enhancing the
security of the aviation industry and passengers has become essential to ensure the
safety and comfort of passengers when travelling by air. For this reason, researchers
have proposed several methodologies such as queueing theory and fuzzy reasoning to
analyse, optimise and study the effects of congestion caused by security screening
processes [1]. And further to strike a balance between increasing security and reducing
processing time, cost and congestion [2].
Queueing theory, first used in 1970, has mainly been applied to optimise the
screening process of a single server. For example, [2] stated that in order to keep the
system stable, the arrival rate must be less than the service rate. [3] aimed to optimise
the security process. They proposed optimal static and dynamic policies for multi-level
security systems. [4] introduced a b (rejection parameter) to determine whether a bag
requires secondary scanning. [5] applied Parkinson’s Law to reduce service time. [6]
introduced a security and accuracy operating characteristic (SAOC). [7] sought to
improve the security screening process by introducing a selectee lane and assigning to
it passengers that may represent a threat.
Different authors used fuzzy reasoning as a method of optimising the security
screening process, as in the work of [8–10], in which linguistic variables and a fuzzy
inference system were used to analyse technical and human factors and enhance a
baggage screening process. The author in [11] presented a fuzzy reasoning system to
improve the efficiency of a baggage screening system with respect to the uncertainty of
machines (X-ray) and human factors.
Most of the aforementioned studies of airport security screening processes that use
a queueing theory as a method have implemented approaches based on a single server,
such as M/M/1 and M/G/1 (M/M/1 was introduced earlier and in M/G/1, where M
stands for a system with Poisson arrivals, G is generally identically distributed, while 1
is the number of servers [12]). This assumption cannot be practically applied for many
reasons, such as differences in human experience, knowledge of the security screening
process and the existence of special needs passengers who require more processing
time, passengers with slow movement and families with more passengers, as well as
differences between machines such as X-ray machines and metal detectors. We con-
sidered these issues in our model during the implementation phase in order to deter-
mine the average waiting time for a whole system with multiple servers. We consider
an M/M/S/K queueing system with Poisson input, exponential service time, multiple
servers and the buffer or area size (where M represents a system with Poisson arrivals
and exponentially distributed service time, S is the number of servers, and K is the
buffer size).
To the best of our knowledge, this study represents the first attempt to explore the
impact of queue formation on optimising the average waiting time.
The aim of this paper is to use a queueing theory as a method to optimise the
security screening process with multi-servers operating in parallel to serve a different
number of passengers during different seasons, such as Christmas, Easter and school
holidays, and time of the day, as this strongly influences the number of passengers. The
implementation of queueing theory requires certain parameters to be predetermined.
Particle Swarm Optimisation (PSO) is used here to optimise the walking time
parameter at different numbers of servers.
The main contributions of this study are summarised as follows:
• We construct a new queue formation model named QQT-PSO based on the Lindley
process (queueing theory) to improve the average waiting time for multi-servers
operating in parallel for security screening areas.
• We incorporate a new “walking time” parameter in our QQT-PSO model to
improve prediction accuracy.
Queue Formation Augmented with Particle Swarm Optimisation 925
Different methods are used in this proposed application to optimise the average waiting
time for the security screening process. The following subsection reviews these
methods.
Cðs; aÞ l
EðWÞ ¼ ð2Þ
ð1 qÞs
C ðs; aÞ l
E ðRÞ ¼ E ðW Þ þ l ¼ þl ð3Þ
ð1 qÞs
926 M. Naji et al.
The Lindley process defined in Eq. (4) is based on a discrete time stochastic
process, which often used to describe the evolution of a queue length over time or to
determine the average waiting time of passengers in a queue. For example, the first
arriving passenger does not need to wait, so Wn ¼ 0. Subsequent passengers will need
to wait if they arrive at a time before the previous passengers are served [14].
We employed Lindley process in this work due to its simplicity and suitability for
security area queue formation. The simplicity is manifested by the applicability of a
single equation for determining the waiting and sojourn time of passengers.
Wn þ 1 ¼ Wn þ Xn IAn ð4Þ
where Wn is the waiting time of the previous passenger, Xn is the service time of the
previous passenger and IAn is the time between the previous and current arrival (where
both passengers are in service or waiting mode). This IAn parameter gives advantage of
Lindley process over standard queuing theory as presented in Eqs. (1) and (2). Figure 1
illustrates this process given Tn is the arrival time of passenger n and Tn þ 1 is the arrival
time of passenger n þ 1.
x i ð t þ 1Þ ¼ x i ð t Þ þ v i ð t þ 1 Þ ð6Þ
where
xi ðtÞ is the particle position for particle i, vi ðtÞ is the particle velocity for particle i, r1 ; r2
are uniform distribution random numbers between 0 and 1, c1 ; c2 are acceleration
coefficients and w is the inertia coefficient to control the impact of the previous velocity
on the updated velocity.
3 QQT-PSO Method
Step 1: To collect the data from the airport and divide the data into two groups. The
first group is to train the PSO model to optimise the Wwalking for different queue
formations. The second group is used to examine the QQT-PSO model to determine the
optimised average waiting time.
Step 2: The aim of PSO is to optimise the objective function known as the fitness
function. In this setting, we follow a wrapper approach based on PSO to seek the global
best walking time that minimises our fitness function defined in Eq. (9). PSO iteratively
tries to opt for the optimal walking time which produces an identical result to the real
data.
The walking time is evaluated based on our fitness function to opt for the best value
that can lead the fitness function fast to zero. A new population of walking time is
generated in each iteration. For each evaluation, PSO optimises a different number of
walking time sets, which is based on the number of servers used during that time.
However, some parameters in PSO have to be predetermined before optimisation and
this includes the following:
The lower bound decision was set at 3 s (aka varmin), while the upper bound of the
decision variable was set at 25 s (aka varmax). The population size was set at 50 and
the number of iterations was set at for 100.
928 M. Naji et al.
Step 3-QQT-PSO Model: In general, most airports use a single passenger queue for
the security screening process. In this study, however, we consider three queuing
scenarios: single, two and three queues. Also, we investigate their impact on passenger
waiting time. The scenarios of our queue formation are presented as follows:
(i) One-queue scenario: all seventeen servers are allocated to one queue area and the
passenger at the top of the queue is allocated to one of the S parallel servers, as
shown in Fig. 3.
(ii) Two-queue scenario: The total area A is divided into two equal queues or buffers
and when there is an even number of servers, each area has the same number,
whereas when there is an odd number of servers, the size of each buffer is based
on the number of servers allocated to it. For example, in the case of seventeen
servers, eight servers are allocated to the first queue and nine servers are allo-
cated to the second queue, as shown in Fig. 4.
(iii) Three-queue scenario: The total area is divided into three queues. The size of
each buffer is based on the number of servers allocated to it. For example, if
there are seventeen servers, the first two queues are allocated six servers each
and the third queue is allocated five servers, as shown in Fig. 5.
PN Qtotal ¼ PN Q1 þ PN Q2 þ . . . þ PN QN ð7Þ
The probability of blocking is defined as the chance that a customer will lose
service due to high demand and lack of resources. For example, a probability of
blocking of 0.01 means that 1% of customers will be denied or will lose service. The
probability of blocking for different numbers of queues for M/M/S/Q (where M rep-
resents a system with Poisson arrivals and exponentially distributed service time, S
stands for the number of servers and Q is the number of queues) is
kl q
Q X
X S 1 kl
r ðrÞ
PQ ¼ qþ1
ð8Þ
q¼1 r¼1 ð1 kl
rÞ
930 M. Naji et al.
Equation (8) illustrates that the probability of blocking decreases when the number
of queues increases.
To be more applicable to queueing processes such as those at an airport security
screening area and to produce accurate results, another parameter such as walking time
must be incorporated in Lindley equation. Thus, the modified formula is
where Wwalking is the time required by a passenger to walk when the passenger comes to
the top of the queue and is the first to be served.
A modified Lindley equation Eq. (9) is used to determine the average waiting time
of N passengers.
According to [4], knowledge of the passenger screening process differs from one
passenger to another. Also, as proposed in [18–20], the efficacy of screening machines
can vary from one to another. Security officers also vary in their experience, which
could affect the time needed to screen luggage (service time). Finally, according to [21]
and in line with the data collected, the time required for a passenger to complete the
security screening process varies between 15 and 60 s for economy passengers (aka
normal lanes), and is 30 s on average, so a time of 30 s is chosen as the mean average
of the exponentially distributed service time. Families and passengers with special
needs (aka slow lanes) take three-and-a-half times longer than normal passengers,
whereas business passengers and the crew (aka express lanes) require half the time
compared to normal passengers. These reasons support our case to apply different
service times for various types of passengers while we optimise the passenger’s
average waiting time. The three service time variations are as follows:
1. Normal lanes: the service time is assumed to be exponentially distributed with
average service mean of 30 s and then multiplied by a random number generator
between 1 and 2. The aim of this variation is to ensure variability among passen-
gers. This is applied to all servers that serve normal passengers, and the service time
of passengers is expressed by the equation:
ðh: lnðRÞÞ ð10Þ
where h is the exponential distributed service time and ðRÞ is random number
generator between 1 and 2.
2. Slow Lanes: various passengers such as families and passengers with special needs
may take three-and-a-half longer than normal passengers. This case is implemented
using the equation:
ððh: lnðRÞÞ 3:5Þ ð11Þ
3. The express or fast lanes: these lanes serve passengers including crew, staff and
business passengers. These passengers may require half the service time of normal
passengers. The fast lane is implemented using the equation:
4.2 Results
The first experiment is to optimise the walking time parameter introduced in Eq. (9) by
applying the PSO algorithm. The inter-arrival rate between two consecutive passengers
is exponentially distributed at the rate of 1/k s. The area size is set to be compatible
with the Sydney airport security area, but the area size cannot be disclosed due to
confidentiality.
In the following sections, we present the results of our QQT-PSO model for one
queue, and then we compare the results with the actual results (airport).
Generally speaking, the aims of this experimental study are twofold. One is to
examine how significant the proposed augmented QQT-PSO model is by producing an
accurate and close result to the airport’s actual system. The second target is to measure
the improvement in the QQT-PSO average waiting time compared with the actual (one
queue) airport result.
number of servers varies between one and three, the QQT-PSO produces an average
waiting between 66% and 88%. The proposed system is able to produce a standard
error (S) prediction interval of 88.16% similar to the actual values across all of the
servers, also considering the variation in service time between passengers, security
officers and machines.
The mean square error shows the error changes between a range 4.77 and 37.17 s
when there are more than three servers. The mean square error increases from 129.66 to
218.54 s when there are three or less than three servers. However, the paired t-test
between the actual results and QQT-PSO produces a p-value of 0.6894, but this result
is not significant at p 0.05. Hence, these results suggest that our result is very
similar to the airport result.
Table 1. Average waiting time and standard error (QT-PSO (1Q, 2Q and 3Q) and actual result)
No. of servers AWT (seconds) MSE (seconds2) S (second) AWT (seconds)
QQT-PSO
Real QQT-PSO Actual-1Q Actual-1Q (2Q) (3Q)
17 389.002 397.406 29.140 5.398 354.307 330.214
16 174.557 180.167 21.183 4.602 163.267 152.168
15 170.535 175.397 19.710 4.439 158.402 147.947
14 256.651 253.112 13.557 3.682 233.384 219.38
13 201.054 196.695 17.742 4.212 178.665 168.302
12 161.045 165.578 18.089 4.253 149.861 141.169
11 160.100 159.545 7.206 2.684 147.229 139.131
10 159.125 162.772 12.980 3.602 150.123 139.131
9 179.370 179.626 4.773 2.184 168.897 160.452
8 150.931 148.600 10.983 3.314 140.415 133.675
7 146.261 140.907 18.994 4.358 135.057 128.709
6 120.291 116.852 14.481 3.805 110.439 105.469
5 98.314 93.915 20.275 4.502 90.810 87.086
4 135.742 129.267 37.174 6.097 128.389 125.821
3 84.228 74.721 129.668 11.387 74.721 74.721
2 77.983 65.972 162.123 14.564 65.972 65.972
1 26.285 39.775 218.549 17.565 39.774 39.774
Average 158.322 165.033 95.933 5.920 146.453 138.730
Figure 6 shows the average waiting time for the proposed model compared to the
actual result from the airport. This figure shows that the QQT-PSO produces very
accurate results compared to the actual result and it also follows the same trend.
Queue Formation Augmented with Particle Swarm Optimisation 933
Fig. 6. Average waiting time (QT-PSO (1Q, 2Q and 3Q)-actual airport result)
5 Conclusion
Airport security screening is essential to ensure the safety of the aviation industry and
passengers travelling by air. Experienced officers and advanced technological instru-
ments help to enhance security. However, the delay caused by security screening
remains a concern. Technical methods have been proposed to model and optimise this
process. Despite the achieved improvements, delays still present a major concern. This
paper used queueing theory to build a system with multi-lane parallel servers to study
the impact of queue formation on the average service time. Unlike existing methods
which do not consider service time variations between the different servers, our pro-
posed method successfully takes this issue into consideration. The results show the
proposed method extensively improved the system average waiting time by 11.98%.
Further studies are required to develop new methods which address the limitations of
the existing methods and the theory behind them by taking into consideration human
and machine factors to determine the optimal numbers of servers required to serve a
different number of passengers to achieve an acceptable waiting time.
Queue Formation Augmented with Particle Swarm Optimisation 935
References
1. Naji, M., et al.: Airport security screening process: a review. In: CICTP 2017.
ASCE LIBRARY (2017)
2. Gilliam, R.R.: An application of queueing theory to airport passenger security screening.
Interfaces 9(4), 117–123 (1979)
3. Lee, A.J., Jacobson, S.H.: The impact of aviation checkpoint queues on optimizing security
screening effectiveness. Reliab. Eng. Syst. Saf. 96(8), 900–911 (2011)
4. Babu, V.L.L., Batta, R., Lin, L.: Passenger grouping under constant threat probability in an
airport security system. Eur. J. Oper. Res. 168(2), 633–644 (2006)
5. Marin, C.V., et al.: Human factors contributes to queuing theory: Parkinson’s law and
security screening. In: Proceedings of the Human Factors and Ergonomics Society Annual
Meeting. SAGE Publications (2007)
6. Olapiriyakul, S., Das, S.: Design and analysis of a two-stage security screening and
inspection system. J. Air Transp. Manage. 13(2), 67–74 (2007)
7. Nie, X., et al.: Simulation-based selectee lane queueing design for passenger checkpoint
screening. Eur. J. Oper. Res. 219(1), 146–155 (2012)
8. Skorupski, J., Uchroński, P.: A fuzzy model for evaluating airport security screeners’ work.
J. Air Transp. Manage. 48, 42–51 (2015)
9. Skorupski, J., Uchroński, P.: Fuzzy inference system for the efficiency assessment of hold
baggage security control at the airport. Saf. Sci. 79, 314–323 (2015)
10. Skorupski, J., Uchroński, P.: A fuzzy system to support the configuration of baggage
screening devices at an airport. Expert Syst. Appl. 44, 114–125 (2016)
11. Skorupski, J., Uchroński, P.: A fuzzy reasoning system for evaluating the efficiency of cabin
baggage screening at airports. Transp. Res. Part C: Emerg. Technol. 54, 157–175 (2015)
12. Cooper, R.B.: Introduction to Queueing Theory. North Holland (1981)
13. Avi-Itzhak, B., Levy, H., Raz, D.: Quantifying fairness in queueing systems: principles and
applications. Preprint (2004)
14. Asmussen, S.: Applied Probability and Queues, vol. 51. Springer, Heidelberg (2008)
15. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks, vol. IV (1995)
16. Braytee, A., et al.: ABC-sampling for balancing imbalanced datasets based on artificial bee
colony algorithm. In: 2015 IEEE 14th International Conference on Machine Learning and
Applications (ICMLA). IEEE (2015)
17. Del Valle, Y., et al.: Particle swarm optimization: basic concepts, variants and applications in
power systems. IEEE Trans. Evol. Comput. 12(2), 171–195 (2008)
18. Wang, X., Zhuang, J.: Balancing congestion and security in the presence of strategic
applicants with private information. Eur. J. Oper. Res. 212(1), 100–111 (2011)
19. Martín-Cejas, R.R.: Tourism service quality begins at the airport. Tour. Manage. 27(5), 874–
877 (2006)
20. Almazroui, S., Wang, W., Zhang, G.: Imaging technologies in aviation security. Adv. Image
Video Process. 3(4), 12 (2015)
21. Kirschenbaum, A.A.: The cost of airport security: the passenger dilemma. J. Air
Transp. Manage. 30, 39–45 (2013)
Polar Topographic Derivatives for 3D
Face Recognition: Application to Internet
of Things Security
1 Introduction
Automatic face recognition has many commercial, military and security appli-
cations such as Internet of Things (IoT) [1,2]. Despite the significant progress in
face recognition technology, recognizing a face under the unconstrained condition
is still challenging. There are five major challenging problems in face recogni-
tion which are changing the illumination condition, facial expressions, head pose
variations, occlusions, and aging [3]. Many of the face recognition algorithms
are categorized as holistic which use the whole face image for feature extraction.
Alyuz et al. [7] constructed seven meaningful regions: the nose, the left/right eye,
the forehead, the left/right cheek, and the mouth-chin. By partitioning the 3D
face image into different paths their scheme was fast and accurate. To heighten
the accuracy, the results were fused in decision level after applying the Lin-
ear Discriminate Analysis (LDA) to patches. Smeets et al. [6] proposed a local
feature method for 3D face, called meshSIFT in which important points are
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 936–945, 2019.
https://doi.org/10.1007/978-3-030-15035-8_92
Polar Topographic Derivatives for 3D Face Recognition 937
The topography map is one of the most important shape representation methods
in 3D space. A topography map shows the physical features of a surface and
their structural relationships. A topography map reveals the relief of a surface
by means of level curves (contour lines). Lines that joint equal-height points on
the surface above or below a reference point is known as a Level curves. The
highest point in the topography map usually known as a reference point. Figure 1
indicates an example of level curves in a face surface and the corresponding
topography map.
Imagine a surface z(x, y), the level curve with the height index h, Lh , is
obtained by
Fig. 1. Level curves and topography map. (a) original 3D face surface, (b) surface level
curves, (c) topography map
where zmax is the maximum height (reference) point in the surface, wz is the step
size of the level curve, h is the height index of the level curve, and · denotes
the floor function.
Given the level curves h , the topography map T (x, y) is indicated as
1, if (x, y, z) ∈ h
T (x, y) = (2)
0, / h
if (x, y, z) ∈
After extracting the topography map from the face surface, Gabor wavelet is
applied to the topography map. The Gabor wavelets are represented as follows
[9]:
kμ,v
2
kμ,v
2
z2 σ2
ϕμ,v (z) = 2 exp − [exp (ikμ,v ) − exp(− )] (3)
σ 2σ 2 2
where μ and v denote the orientation and the scale of the Gabor wavelets,
respectively. In Eq. (3), kμ,v is defined as
2 2
r= (x − xrp ) + (y − yrp ) (6)
Polar Topographic Derivatives for 3D Face Recognition 939
⎧
⎪
⎨ 0, if x = xrp , y = yrp
−1 y−yrp
θ= tan ( x−xrp ), if x ≥ xrp (7)
⎪
⎩ tan−1 ( y−yrp ) + π,
x−xrp if x < xrp
where xrp and yrp are the coordinates of the reference point (nose tip). Polar
Gabor topography maps are divided into non-overlapping polar patches. Here,
we label the patch points with p and q indexes as
the patch sizes in the radial and the angular directions are shown by wr and wθ
respectively, and indicate as
wθ = 2π/3(2p − 1) (11)
where rmax and Nr indicate the maximum radius and the number of sectors in
the radial direction, respectively.
In the proposed polar patching, the total amount of patches, Np , is
rmax /wr
Np = 3(2k − 1) (12)
k=1
The (p, q)th patch of the Gabor topography map, GTpq (rpq , θpq ), is repre-
sented as
pq
P GTμ,v (rpq , θpq ) = P GTμ,v (wr (p − 1) + rpq , wθ (q − 1) + θpq ) (13)
where rpq and θpq are the coordinates of the (p, q)th patch. For each patch,
we compute Polar Average Absolute Deviation (PAADs) from the mean of the
patch’s gray levels. For the (p, q)th patch of the polar Gabor topography map,
the PAAD feature from the mean of the gray levels is defined as
pq 1
pq pq pq pq pq pq
P AADμ,v = P GTμ,v (r , θ ) − P GTμ,v (r , θ ) (14)
N rpq pq
θ
pq pq pq
where P GTμ,v (r , θ ) is the (p, q)th patch of the Polar Gabor topography map.
pq
N is the number of pixels in the (p, q)th patch and P GTμ,v (rpq , θpq ) is the mean
of the (p, q)th patch’s gray levels.
By concatenating the PAADs of (p, q)th patch in different scales and orien-
tations, we obtain a new feature vector, called Polar Gabor Topography Vector
(PGTV), as
P GT V pq = P AADμ,v pq
|μ = 0, . . . , 7 & v = 0, . . . , 4 (15)
940 F. Hajati et al.
pq
where P AADμ,v is computed using the μ and v as the scale and the orientation.
We also use high-order polar information in the texture map of the object to
extract discriminant features. First, we transform the texture map into the polar
coordinates. Given a texture map, f (x, y), the transformed texture map in the
polar coordinates, F (r, θ), is defined as
where rpq and θpq are the coordinates of the (p, q)th patch of the partitioned
texture map.
Let F pq (r0pq , θ0pq ) be a point of the (p, q)th patch of the texture map, the
eight neighbors of the point are defined as
where Δr and Δθ are the radial and the angular resolutions of the texture map.
Here, we assume Δθ = π · Δr/180.
The angular directions with different orders and the derivative along the
radial can be computed after determining the 8-neighborhood of the point
F pq (r0pq , θ0pq ). The first-order polar derivatives along the radial and the angu-
lar directions of F pq (r0pq , θ0pq ) are defined as
∂ 2 F pq (r0pq , θ0pq ) /∂r2 = ∂F pq (r0pq , θ0pq ) /∂r − ∂F pq (r1pq , θ0pq )/∂r (21)
∂ 2 F pq (r0pq , θ0pq ) /∂θ2 = ∂F pq (r0pq , θ0pq ) /∂θ − ∂F pq (r0pq , θ1pq )/∂θ (22)
Generally, the nth -order polar derivatives are defined using the (n−1)th -order
polar derivatives as
By using the unit step function, the computed derivatives are encoded
into binary bits, there for binary patterns called Polar Derivative Patterns
(PDP) are formed. The nth -order Polar Derivative Patterns (PDP) along
the radial direction, P DPr (F pq (r0pq , θ0pq )), and along the angular direction,
P DPθ (F pq (r0pq , θ0pq )), in the (p, q)th patch are defined as
P DPrpq (F pq (r0pq , θ0pq )) =
20 .u ∂ (n) F pq (r0pq , θ0pq )/∂r(n) · ∂ (n) F pq (r1pq , θ1pq )/∂r(n) + · · · (25)
+27 .u ∂ (n) F pq (r0pq , θ0pq )/∂r(n) · ∂ (n) F pq (r0pq , θ1pq )/∂r(n)
3 Experimental Results
The robustness of the proposed algorithm in the face recognition task under the
pose rotation and the expression variation is investigated. In our experiments,
the Bosphorus database [14] is used that contains faces with pose rotations
and expression changes. In the experiments, preprocessing was applied to face
scans in the database. Spikes in 3D faces were removed using median filtering in
the z-coordinate. Holes were filled using cubic interpolation and 3D scans were
resampled on a uniform square grid at 1 mm resolution. Finally, all scans were
subjected to pose normalization and cropped to the size of 120 × 120 pixels.
The Bosphorus database [14] contains 105 subjects (61 females and 44 males)
with a total of 4666 faces. The images include frontal views with different
facial expressions, pose variations, and occlusions. The resolution in x, y, and
z (height) dimensions are 0.3 mm, 0.3 mm, and 0.4 mm respectively, and color
texture images are high resolution (1600×1200 pixels). To compute the accuracy
of the proposed method, we consider faces with expression and pose variations
in the database (see Fig. 2). The robustness against the pose rotation is one
of the most major problems in the accuracy computing of the face recognition
systems. Here, we evaluate the acceptable performance of the proposed method
using faces with pose variations from the Bosphorus database [14]. There are 11
different poses including left 45◦ , right 10◦ , right 20◦ , right 30◦ , right 45◦ , right
down, right up, slightly up, strong up, slightly down, and strong down for each
subject. Since our goal is to solve the problem of the face recognition using one
exemplar per person case, as the gallery, one frontal image per enrolled subject
is used.
The rank-1 recognition rate of the our approach is shown in Table 1. We
compare proposed method with the recognition rates of the methods in [8] and
[10] as the benchmarks. For faces with small yaw rotations (right 10◦ , right
20◦ , and right 30◦ ), the rank-1 recognition rate is higher than 98%, while it is
higher than 91% for faces with large yaw rotations (right 45◦ and left 45◦ ). In
upward pitch rotations, our method achieves 100% recognition rate for small and
Fig. 2. Images of a subject in the Bosphorus database [14] with different poses and
expressions.
Polar Topographic Derivatives for 3D Face Recognition 943
Table 1. Comparison the rank-1 recognition rates of the proposed method and the
benchmarks under the pose variations.
strong rotations, while the benchmarks have 87.6% and 81.9% for small rota-
tions and 79% and 69.5% for strong rotations, respectively. For downward pitch
rotations, the accuracy rate of the proposed method is 99.1% and 97.1% in small
and strong rotations, respectively. The rank-1 recognition rate of the proposed
method is 94.3% and 90.5% for faces with the right-upward rotation and the
right-downward rotation, respectively. Overally, the proposed method achieves
the average recognition rate of 96.5% across all rotations in the database, out-
performing [8] and [10] with 69.1% and 50%, respectively.
We also evaluate the performance of the proposed method under the expres-
sion changes using the Bosphorus database [14]. For this purpose, we report the
rank-1 recognition rate of the proposed method using all faces with pose and
Table 2. Comparison the rank-1 recognition rates of the proposed method and the
benchmarks under the expression variations.
expression variations as the probe (see Table 2). The rank-1 recognition rate for
the proposed method is 96.6% using 105 galleries and 3696 probes, while the
best one among the benchmarks is 95.3% using 47 galleries and 1508 probes.
4 Conclusion
In this paper, a new approach called Polar Topographic Derivatives (PTD) was
proposed for 3D face recognition. The PTD used both the range and the tex-
ture maps to extract discriminative features. We applied the PTD to expression
and pose invariant face recognition purpose, and achieved considerable improve-
ments in the recognition rates compared to the state-of-the-art methods. The
experimental results demonstrate that the algorithm has an acceptable perfor-
mance even though in large database. Although the proposed algorithm have
been applied to 3D face recognition task, it can be generalized to other textured
3D shapes.
References
1. Hajati, F., Tavakolian, M., Gheisari, S., Gao, Y., Mian, A.S.: Dynamic texture
comparison using derivative sparse representation: application to video-based face
recognition. IEEE Trans. Hum.-Mach. Syst. 47(6), 970–982 (2017)
2. Hajati, F., Cheraghian, A., Gheisari, S., Gao, Y., Mian, A.S.: Surface geodesic
pattern for 3D deformable texture matching. Pattern Recognit. 62, 21–32 (2017)
3. Abate, A.F., Nappi, M., Riccio, D., Sabatino, G.: 2d and 3d face recognition: a
survey. Pattern Recognit. Lett. 28(14), 1885–1906 (2007)
4. Chang, K.I., Bowyer, K.W., Flynn, P.J.: Face recognition using 2d and 3d facial
data. In: ACM Workshop on Multimodal User Authentication, pp. 25–32 (2003)
5. Bowyer, K.W., Chang, K., Flynn, P.: A survey of approaches and challenges in 3d
and multi-modal 3d + 2d face recognition. Comput. Vis. Image Underst. 101(1),
1–15 (2006)
6. Smeets, D., Keustermans, J., Vandermeulen, D., Suetens, P.: meshsift: Local sur-
face features for 3d face recognition under expression variations and partial data.
Comput. Vis. Image Underst. 117(2), 158–169 (2013)
7. Alyuz, N., Gokberk, B., Akarun, L.: Regional registration for expression resistant
3-d face recognition. IEEE Trans Forensics Secur. 5(3), 425–440 (2010)
8. Hajati, F., Raie, A.A., Gao, Y.: 2.5d face recognition using patch geodesic
moments. Pattern Recognit. 45(3), 969–982 (2012)
9. Lee, T.: Image representation using 2d gabor wavelets. IEEE Trans Pattern Anal.
Mach. Intell. 18(10), 959–971 (1996)
10. Malassiotis, S., Strintzis, M.G.: Robust face recognition using 2d and 3d data: pose
and illumination compensation. Pattern Recognit. 38(12), 2537–2548 (2005)
11. Bronstein, A.M., Bronstein, M.M., Kimmel, R.: Three-dimensional face recogni-
tion. Int. J. Comput. Vis. 64(1), 5–30 (2005)
12. Mian, A.S., Bennamoun, M., Owens, R.: An efficient multimodal 2d–3d hybrid
approach to automatic face recognition. IEEE Trans. Pattern Anal. Mach. Intell.
29(11), 1927–1943 (2007)
Polar Topographic Derivatives for 3D Face Recognition 945
13. Zhang, H., Gao, W., Chen, X., Zhao, D.: Learning informative features for spatial
histogram-based object detection. In: Proceedings of the International Conference
on Neural Networks, vol. 3, pp. 1806–1811 (2005)
14. Savran, A., Alyüz, N., Dibeklioğlu, H., Çeliktutan, O., Gökberk, B., Sankur, B.,
Akarun, L.: Biometrics and identity management. Bosphorus Database for 3D Face
Analysis, pp. 47–56 (2008)
15. Alyüz, N., Gökberk, B., Dibeklioğlu, H., Savran, A., Salah, A., Akarun, L., Sankur,
B.: Biometrics and identity management. In: Biometrics and Identity Management,
pp. 57–66. Springer, Heidelberg (2008)
16. Alyüz, N., Gökberk, B., Akarun, L.: A 3d face recognition system for expression and
occlusion invariance. In: Proceedings of the IEEE Second International Conference
on Biometrics Theory, Applications and Systems, pp. 1–7 (2008)
17. Dibeklioğlu, H., Gökberk, B., Akarun, L.: Nasal region-based 3d face recognition
under pose and expression variations. In: Proceedings of the Third International
Conference on Advances in Biometrics, vol. 5558, pp. 309–318 (2009)
A Survey on Conversational Agents/Chatbots
Classification and Design Techniques
1 Introduction
The interaction between humans and computer systems is shifting towards natural
language-based interfaces. Natural language has recently been viewed as a compelling
enabling technology for personalization, which allows each user to interact with the
system in his or her own words, rather than using one of a small number of preset ways
to interact [1]. A text based, or a voice-based conversation is usually initiated by the
user who asks a question using natural language and the chatbot answers the question
using natural language [2]. Chatbots are primarily designed to conduct a conversation
with humans using natural language dialogues via auditory or textual methods, in such
a way that humans can easily communicate with the chatbots.
Chatbots have already come a long way from their roots in systems that were more
about simple chat or fun and entertainment [2]. Although modern chatbots use the
power of artificial intelligence to answer users’ complex questions but current state-of-
the art systems are still a long way from being able to have coherent, contextual and
natural conversations with humans [3].
Chatbot technology is not new; it was started in 1960’s when Joseph Weizen-
baum’s program ELIZA, published in 1966, which seemed to be able to fool users into
believing that they were conversing with a real human [4]. Early chatbots like ELIZA
used keyword matching and minimal context identification and lacked the ability to
keep a conversation going. ELIZA program inspects the user input for keywords and
once a keyword is found, it transformed the sentence according to the rules associated
with that keyword in a script. Response is generated by assembly/reassembly rules
associated with selected decomposition rules. Keywords and their associated trans-
formation rules constitute the SCRIPT (it is data and not a part of the program itself) for
a particular class of conversation [5]. A.L.I.C.E. (Artificial Linguistic Internet Com-
puter Entity) is another natural language processing chatbot and is one of the strongest
programs of its type. It has won the Loebner Prize, awarded to accomplished huma-
noid, talking robots, three times (in 2000, 2001 and 2004) [6]. A.L.I.C.E. is built using
Artificial Intelligence Mark-up Language (AIML), based on eXtensible Markup Lan-
guage (XML), developed over the past 10 years [7]. AIML supports most of the
modern chatbot platforms in use today. AIML based chatbot are the most popular
because they are lightweight and easy to configure. Predefined responses for a given set
of questions also induces constraint on the conversational agent, in the sense that
appropriate context also needs to be considered while generating these responses.
Developers and researchers continue to add new components with improved func-
tionalities to the existing methodologies and in some cases, have even proposed a new
design architecture. Most of these improvements were largely based on utilization of
ontologies and recollecting details from the current and previous conversation [8].
Chatbots are generally designed either to serve some specific task, also called task-
oriented chatbots or to a simple chit-chat, also called non-task oriented chatbots, for fun
and entertainment. In this paper, the broad and main classification of chatbots is pre-
sented first which is followed by the exploration of popular chatbot design approaches
used in the categories mentioned above in this paper.
2 Classification of Chatbots
Over the last few years, the chatbot field has become so dynamic with the arrival of new
technologies that a precise classification of chatbots has become subjective to the scope
of their use. Chatbots could be classified into various categories based on several criteria,
e.g. mode of interaction, knowledge domain, their usage and the design techniques
(response generation method) that are typically employed in building these chatbots.
These criteria may include the core design philosophy of the chatbots or the extent to
which context needs to be stored and considered in understanding the conversation or the
type and purpose of the conversation for which the chatbot needs to be designed [8]. The
broad classification can be done based on the following criteria (Fig. 1);
1. Interaction Mode (Text based, or voice/speech based)
2. Chatbot Application (Task-Oriented or Non-Task-Oriented)
3. Rule-based or AI (Machine learning, Deep learning etc.)
4. Domain-Specific or Open-Domain
948 S. Hussain et al.
But in general, chatbots are classified into two main categories based on goals
[9–11] (Fig. 2):
1. Task-oriented chatbots
2. Non-task-oriented chatbots
Task-oriented chatbots are designed for a particular task and are set up to have short
conversations, usually within a closed domain. Unlike task-oriented chatbots, non-task
oriented chatbots can simulate a conversation with a person and seem to perform chit-
chat for entertainment purpose in open domains [10].
There are several different approaches that can be used to develop a task-oriented or
non-task oriented chatbot. These approaches can overlap between these two main
categories of chatbots. This section explores those approaches, in a broad scope, used
A Survey on Conversational Agents/Chatbots Classification and Design Techniques 949
in the development of chatbots since the very early chatbot ELIZA. These approaches
can be divided into three main categories [8, 10, 12–15];
1. Rule-Based Approaches
2. Retrieval-Based Approaches
3. Generative-Based Approaches
Each of these main approaches can employ following multiple techniques;
1. Parsing
2. Pattern Matching
3. AIML
4. Chatscript
5. Ontologies
6. Markov Chain Model
7. Artificial Neural Networks Models
7:1 Recurrent Neural Networks (RNNs)
7:2 Sequence to Sequence Neural Model
7:3 Long Short-Term Memory Networks (LSTMs)
1. Parsing. Parsing is a method that takes the text as an input and extract meaningful
information from it by converting that text into a set of more uncomplicated words
(lexical parsing) that can be easily stored and manipulated. It determines the
grammatical structure of a sentence. A simplistic parsing technique was used in an
early chatbot, ELIZA to parse the input text for the keyword in the sentence. The
keyword is then matched against the documents in the corpus to find the appropriate
response to the user [16]. A more advanced parsing technique is semantic parsing,
which converts user’s input sentence to a machine-understandable representation of
its meaning. It is used in some of the modern commercial chatbots tools like
Dialogflow to understand the user’s intent [17].
2. Pattern Matching. Pattern matching is the most commonly used approach in chat-
bots to classify the user input as <pattern> and produce a suitable response stored
in <template>. These <pattern> <template> pairs are hand crafted. Although pattern
matching techniques are used in both early and modern chatbots, the complexity of the
algorithms used in them differs. Early chatbots like ELIZA used simple pattern
matching rules, as compared to A.L.I.C.E, which uses more complex pattern matching
rules and associates some degree of conversational context while searching the stored
categories for a matching <pattern> [7]. Pattern matching technique is mostly used in
QA chatbots. This techniques has advantage of being flexible to create conversations.
The approach has some disadvantages in terms of scaling as all the possible patterns
are built manually and it’s not a trivial task. Due to the scaling issue, chatbots infor-
mation extraction capabilities are limited, and its responses could be predictable and
repetitive.
3. AIML. AIML (Artificial Intelligence Mark-up Language) has been derived from
XML (Extensible Mark-up Language). AIML is designed to create conversational
flow in chatbots. AIML is an open standard. AIML is made up of data objects which
are called AIML objects. These AIML objects, which are also called AIML
950 S. Hussain et al.
elements, are made up of units which are called topics and categories. The topic is
an optional top-level element that has a name attribute and a set of categories related
to that topic. Categories are the most fundamental unit of knowledge in an AIML
file. A category has a minimum of two more elements in it called ‘pattern’ and
‘template’. Each category is a rule for matching an input and converting it to an
output. The ‘pattern’ matches against the user input and the ‘template’ is used to
generate the Chatbot’s answer [18]. AIML is powerful in designing chatbots’
conversational flow, flexible and easy to use but needs some Natural language
processing and programming expertise.
4. Chatscript. Chatscript is an open source authoring tool for building chatbots.
Chatscript is a combination of natural language engine and dialog management
system designed for interactive conversation while maintaining user state across the
conversations [19]. It is a rule-based engine where rules are created in program scripts
through a process called dialog flow scripting. These scripts can be stored as a normal
text file. Machine learning tools can also be used to mine users conversations logs to
improve your dialog flows. Chatscript make use of concepts. A concept of all nouns
or adverbs can be created. An existing database of some 2000 predefined concepts
and scripters can also easily write their own concepts [20].
5. Ontologies. Domain ontologies are used in chatbots to replace hand crafted domain
knowledge with ontological domain knowledge. Use of ontologies is not new as
they have been used within specific dialogue system modules, for example in
language generation as the basis of the systemic grammar approach [21]. The main
advantage of using domain ontology in a chatbot is that the chatbot can explore the
concepts nodes of an ontology to establish the relationship between concepts that
are used in conversation but can also imply new reasoning [22]. However more
general use of ontologies in chatbots is relatively small.
6. Markov Chain Model. A Markov chain model is a probabilistic model and the
rules in the Markov chain process are based on probabilities. The Markov chain
seeks to model probabilities of state transitions over time. The key idea behind
Markov chain model is that given the current state, there is a fixed probability for
one or more states that it will go to next. When a chatbot uses this method, it will
produce an output that goes with the state transition. This allows the chatbot to
construct sentences for responses that are more suitable probabilistically whilst
being different every time but are more or less coherent. The starting state can be
based on the user input, giving the response some relevance. Markov Chains are a
popular method to create chatbots for entertainment purposes that imitate simple
human conversation. Because Markov chain model is a simplified version of a
complex decision-making process, it does not work well to emulate rich and
complex conversation [8]. A simple Markov chains model is easy to program and
the whole model can be summarized in a matrix.
7. Artificial Neural Networks Models. The latest advances in machine learning,
specifically in artificial neural networks, has made it possible to develop more
intelligent chatbots. Artificial Neural Networks based chatbots can use both
retrieval-based or generative-based approaches in terms of response generation,
although the research trend is moving into the generative-based approach [23, 24].
The main difference between rule-based and artificial neural network based chatbot
A Survey on Conversational Agents/Chatbots Classification and Design Techniques 951
Sequence model (Seq2Seq) is the current industry best practice for response
generation and is widely used [29].
7:3 Long Short-Term Memory Networks (LSTMs). Long Short-Term memory
networks (LSTMs) are a special kind of recurrent neural networks [8]. LSTMs
are designed to avoid the long term dependency problem of RNNs. Memory
cells and gates are introduced in LSTMs which allows the cells to remember the
previous information for long periods of time. These memory cells can store
information in, write new information and read information from them like data
in a computer’s memory. The flow of information is controlled with the use of
input gates, forget gates and output gates [30]. An LSTM network is ideal for
learning from experience as compared to a traditional RNN. LSTM networks
have since replaced RNNs as standard for this task. A well-trained LSTM
network has the capability to perform better classification, processing and
prediction of time series even when there is a long period of gaps of unknown
size between significant events. These capabilities exemplify LSTM’s high
performance as compared to other available RNNs, hidden Markov models, and
other sequence learning methods that are being used in many applications. Thus
LSTMs prove to be highly useful in designing chatbots due to their ability to
refer to a piece of distant information in time very frequently [8, 30].
4 Task-Oriented Chatbots
The main aim for task-oriented chatbots are to help the user to achieve a certain task.
They are designed for dealing with specific scenarios such as: booking a hotel/flight,
booking accommodations, placing an order for a product, scheduling an event, or
helping users to access some specific information etc. Personal assistants such as
Cortana, Alexa and Siri are examples of voice based task-oriented chatbots/
conversational agents who attempt to return an answer for the task they receive.
Task-oriented chatbots work well in restricted domains. These chatbots do not possess
general knowledge and you cannot ask them trivia questions. Rather, they are goal
oriented chatbots focused on helping you achieve a specific goal.
5 Non-task-Oriented Chatbots
Non-task oriented chatbots are systems designed for extended conversations, set up to
mimic the unstructured conversational or ‘chats’ characteristic of human-human
interaction, rather than focusing on a particular task like booking plane flights. These
systems often have an entertainment value. Non-task oriented chatbot architectures fall
into two classes:
1. Generative-Based such as sequence-to-sequence models, which generate proper
responses during the conversation
2. Retrieval-Based which learn to select responses from the current conversation from
a repository.
A Survey on Conversational Agents/Chatbots Classification and Design Techniques 953
Generative models can generate more proper responses that could have never
appeared in the corpus. Corpus-based systems mine large datasets of human-human
conversations, which can be done by using information retrieval (IR-based systems
simply copy a human’s response from a previous conversation) or by using a machine
translation paradigm such as neural network sequence-to- sequence systems, to learn to
map from a user utterance to a system response. Retrieval-based models enjoy the
advantage of informative and fluent responses [31], because they select a proper
response for the current conversation from a repository with response selection algo-
rithms. A variety of heuristics can be applied in order to choose a proper response. It
may be based on the fairly simple concept of a rule-based expression match or may
well be as complex as using a combination of Machine Learning classifiers. Systems
based on this concept do not produce new responses but just choose one from a pool of
predefined responses. Rule-based systems include the early influential ELIZA and
PARRY systems.
Non-task oriented chatbots typically do not have any goal to achieve and are not
designed to handle any particular task. Nonetheless, these chatbots are very helpful in
different applications, such as keeping elderly people company and improving the
understanding of the learners of the second language. In each of its applications, unlike
task-oriented chatbots, they might not have to observe a set of similar tasks; rather, they
would be required to respond to the users in order to proceed on the discussion [35].
Therefore, the algorithm designing in case of non-task oriented chatbots will be
954 S. Hussain et al.
complex as compared to a task-oriented chatbot. In some cases, these chatbots are also
combined with task-oriented chatbots in order to provide an enhanced experience to the
users for a smooth transition. The dialogue context in this case will be random and the
user might not have discussion related to any particular context [36]. For instance, the
user may ask the chatbot for information related to weather, or the question might be
related to horoscope or any recent news story. Therefore, the chatbot must be capable
of retrieving information from online resources in order to communicate effectively
with the user.
8 Conclusion
In this paper, we have explored chatbot’s broad and general classification and discussed
various design techniques used to design chatbots. To conclude, it can be argued that
the typical interaction between human and computer systems has been modernised and
natural languages have become a major input in contemporary chatbots. Natural lan-
guage has recently been viewed as a compelling enabling technology for personal-
ization, which allows each user to interact with the system in his or her own words,
rather than using one of a small number of preset ways to interact. In handling such
queries based on the natural language, the chatbots are programmed using different
strategies and algorithms in order to generate an effective response for the user. In
earlier discussion on the topics, the chatbots can be typically divided into two main
categories including task-oriented chatbots and non-task oriented chatbots. As the name
highlights, the task-oriented chatbots are designed to perform particular tasks based on
the instructions of the user while a non-task oriented chatbot will have several purposes
but will not be able to perform any specific task. The dialogue context in both of these
chatbots will be different. For instance, in case of a task-oriented chatbot, it might have
to process a set of similar inputs and simple algorithms based on pattern matching can
be used to process the queries. In case of non-task oriented chatbot, the queries may
vary in terms of their context and several algorithms and strategies would be required to
use in order to assure correct response of the system. Generative-based chatbots are
difficult to build and operate as compared to retrieval-based chatbots.
References
1. Zadrozny, W., et al.: Natural language dialogue for personalized interaction. Commun. ACM
43(8), 116–120 (2000)
2. Hussain, S., Athula, G.: Extending a conventional chatbot knowledge base to external
knowledge source and introducing user based sessions for diabetes education. In: 2018 32nd
International Conference on Advanced Information Networking and Applications Work-
shops (WAINA) 2018. IEEE
3. Levesque, H.J.: Common sense, the turing test, and the quest for real AI: Reflections on
Natural and Artificial Intelligence. MIT Press, Cambridge (2017)
4. Wikipedia: Chatbot. 29 December 2018 (cited 30 December 2018). https://en.wikipedia.org/
wiki/Chatbot
A Survey on Conversational Agents/Chatbots Classification and Design Techniques 955
5. Weizenbaum, J.: ELIZA–a computer program for the study of natural language commu-
nication between man and machine. Commun. ACM 26(1), 23–28 (1983)
6. Wikipedia. Artificial Linguistic Internet Computer Entity. 19 November 2018 (cited 30
December 2018). https://en.wikipedia.org/wiki/Artificial_Linguistic_Internet_Computer_
Entity
7. Wallace, R.S.: The anatomy of ALICE. In: Parsing the Turing Test, pp. 181–210. Springer,
Dordrecht (2009)
8. Ramesh, K., et al.: A Survey of Design Techniques for Conversational Agents. In:
International Conference on Information, Communication and Computing Technology.
Springer, Singapore (2017)
9. Budulan, S.: Chatbot Categories and Their Limitations (2018). (cited May 2018). https://
dzone.com/articles/chatbots-categories-and-their-limitations-1
10. Chen, H., et al.: A Survey on Dialogue Systems: Recent Advances and New Frontiers
(2017). arXiv preprint arXiv:1711.01731
11. Yan, Z., et al.: Building Task-Oriented Dialogue Systems for Online Shopping. In: AAAI
(2017)
12. Masche, J., Le, N.-T.: A Review of Technologies for Conversational Systems. In:
International Conference on Computer Science, Applied Mathematics and Applications.
Springer, Cham (2017)
13. Abdul-Kader, S.A., Woods, J.: Survey on chatbot design techniques in speech conversation
systems. Int. J. Adv. Comput. Sci. Appl. 6(7), 72–80 (2015)
14. Mathur, V., Singh, A.: The Rapidly Changing Landscape of Conversational Agents (2018).
arXiv preprint arXiv:1803.08419
15. Bradeško, L., Mladenić, D.: A survey of chatbot systems through a loebner prize
competition. In: Proceedings of Slovenian Language Technologies Society Eighth Confer-
ence of Language Technologies (2012)
16. Weizenbaum, J.: ELIZA—a computer program for the study of natural language
communication between man and machine. Commun. ACM 9(1), 36–45 (1966)
17. Clarke, D.: Three AI Technologies that could make chatbots intelligent. 24 March 2018
(cited 18 December 2018). https://hackernoon.com/three-ai-technologies-that-could-make-
chatbots-intelligent-10f8c6e8b4b0
18. Shawar, B.A., Atwell, E.: Chatbots: are they really useful? Zeitschrift für Computerlinguistik
und Sprachtechnologie, p. 29 (2007)
19. Wikipedia contributors. ChatScript. Wikipedia, The Free Encyclopedia. 4 September 2018,
19:19 UTC. https://en.wikipedia.org/w/index.php?title=ChatScript&oldid=858055386. Acce-
ssed 22 Jan 2019
20. Robino, G.: How to build your first chatbot using Chatscript (2018). (cited 18 December
2018). https://medium.freecodecamp.org/chatscript-for-beginners-chatbots-developers-c58b
b591da8
21. Milward, D., Beveridge, M.: Ontology-based dialogue systems. In: Proceedings of 3rd
Workshop on Knowledge and reasoning in practical dialogue systems (IJCAI 2003) 2003
22. Al-Zubaide, H., Issa, A.A.: Ontbot: Ontology based chatbot. In: 2011 Fourth International
Symposium on Innovation in Information & Communication Technology (ISIICT). IEEE
(2011)
23. Nuez Ezquerra, A.: Implementing ChatBots using Neural Machine Translation Techniques
(2018). Universitat Politècnica de Catalunya
24. Csáky, R.: Deep Learning Based Chatbot Models (2017). https://doi.org/10.13140/rg.2.2.
21857.40801
25. Pelk, H.: Machine learning, neural networks and algorithms (2016). (cited 18 January 2019).
https://itnext.io/machine-learning-neural-networks-and-algorithms-4cd27a105b89
956 S. Hussain et al.
26. Wikipedia contributors. Artificial neural network. In: Wikipedia, The Free Encyclopedia.
https://en.wikipedia.org/w/index.php?title=Artificial_neural_network&oldid=878429976.
Accessed 13:24, 25 Jan 2019
27. Young, T., et al.: Recent trends in deep learning based natural language processing. IEEE
Comput. Intell. Mag. 13(3), 55–75 (2018)
28. Cahn, J.: CHATBOT: Architecture, design, and development. In: University of Pennsylvania
School of Engineering and Applied Science Department of Computer and Information
Science (2017)
29. Cho, K., et al.: Learning phrase representations using RNN encoder-decoder for statistical
machine translation (2014). arXiv preprint arXiv:1406.1078
30. Sak, H., Senior, A., Beaufays, F.: Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. In: Fifteenth Annual Conference of the
International Speech Communication Association (2014)
31. Ji, Z., et al.: An information retrieval approach to short text conversation (2014). arXiv
preprint. arXiv:1408.6988.1
32. Duijst, D., Sandberg, J., Buzzo, D.: Can we Improve the User Experience of Chatbots with
Personalisation? (2017)
33. Liu, B., Xu, Z., Sun, C., Wang, B., Wang, X., Wong, D.F., Zhang, M.: Content-oriented user
modeling for personalized response ranking in Chatbots. IEEE/ACM Trans. Audio Speech
Lang. Process. (2018)
34. Pilato, G., Augello, A., Gaglio, S.: A modular architecture for adaptive ChatBots. In:
Proceedings of 5th IEEE International Conference on Semantic Computing, ICSC 2011
(2011)
35. Vtyurina, A., Savenkov, D., Agichtein, E., Clarke, C.L.A.: Exploring conversational search
with humans, assistants, and wizards. In: Proceedings of the 2017 CHI Conference Extended
Abstracts on Human Factors in Computing Systems – CHI EA 2017 (2017)
36. Niculescu, A.I., Banchs, R.E.: Strategies to cope with errors in human-machine spoken
interactions: using chatbots as back-off mechanism for task-oriented dialogues. In: Errors by
Humans Mach. multimedia, multimodal Multiling, Data Process (2015)
Dimensionality Reduction for Network
Anomalies Detection: A Deep
Learning Approach
1 Introduction
The deep learning has been used successfully for several applications. In image
recognition, Hinton et al. utilized Deep Belief Network (DNN) on MNIST a numbers
handwritten images dataset, the algorithm recorded 1.25% error rate while the second
lowest was 1.4% scored by SVM [4]. In 2012, ImageNet competition the convolutional
neural network decreased the error rate of 16.4%; the dataset included 14 m labelled
images and 1k categories [6]. In voice recognition. Mohamed et al. utilized DBN to
reduce the phoneme error rate from 26% to 20.7% in TIMIT dataset for voice
recognition [7].
An inclusive study of unsupervised machine learning algorithms evaluated seven
algorithms [8]. The study used classic KDD99 dataset. It deduced all algorithms per-
formed poorly in remote code execution escalation attacks. SVM and Y-means scored
Dimensionality Reduction for Network Anomalies Detection 959
The first stage of the suggested framework is based on the idea of dimensionality
reduction. The concept of dimensionality reduction refers to projecting highly
dimensioned data on a lower subspace without a significant loss of data meaning.
Additionally, in a lower dimensional data, the discrimination between normal and
abnormal classes is evident. Principal Components Analysis PCA is a dimensionality
reduction algorithm learns linear relationships. Kernel PCA is a non-linear version of
the basic PCA. To represent nonlinear relationships, a kernel function is used to map
the data to higher dimensions, and then use PCA to reduce the dimensionality.
Autoencoders algorithms imply dimensionality reduction concepts, as it converts
the data into a new representation that keep most significant features (encoder), and
then execute a reconstruction phase (decoder).
Figure 2 depicts the basic structure of an autoencoder includes two symmetric input
and output layers, in between a series of hidden layers with a lower number of features
(neurons), reducing the no of features is a reduction of dimensionality. However, at the
output layers, autoencoders reconstruct the same number of input features.
V1
Weights
The figure depicts the main components of the proposed detection system. The
framework is established on unsupervised deep neural networks. Two types of USDL
algorithms were used Autoencoders and Restricted Boltzmann machines.
In the context of networks anomalies, threats are continuously growing. Hence,
from a security perspective, an intrusion detection system must be capable of detecting
attacks which have never seen before. This concept is theoretically achievable through
unsupervised learning. Where the supervised algorithms-whether they are classic
algorithms or deep learning must previously see samples and classes to classify the new
records (Fig. 4).
962 A. Dawoud et al.
ReconstrucƟ A
on Error
N
K-
Mea N
ns
A
N
The simulation is conducted in two main scenarios based on AE. Figure 5 shows the
main scenario is an implementation for our algorithm based on AE. The sub-scenarios
use two different simple classical algorithms K-means and Shift- means algorithms at
the second phase for clustering. The scenarios occur as following,
• Autoencoder phase then K-means for clustering
• Autoencoder phase then Shift-means for clustering
We used Tensorflow (TF) as a deep learning development library. Tensorflow is
matrices flow in a graph model. TF graph consists of nodes and edges; nodes represent
mathematical operations, edges represent multi-dimensional data arrays (tensors).
Dimensionality Reduction for Network Anomalies Detection 963
KDD99 is the most used dataset in machine learning and intrusion detection. The
dataset represents real network traffic collected data. The dataset includes 4898431
traffic records for the training, 311029 records for testing [15].
To summarize the results of the simulation, we used confusion matrix statistics. The
confusion matrix consists of columns and rows that list the number of testing samples
as either predicted or actual ratios. Figure 6 is a general description of the confusion
matrix, where we have two classes normal and abnormal. To validate the results the
second phase was done in two different clustering algorithms K-means and shift means.
The confusion matrices for both with different samples sizes are listed Fig. 7. Table 1
shows the AE scored higher than 99% accuracy with K-means
The accuracy = (TP + TN) / (P + N), from Fig. 7 we conclude Table 1 to sum-
marize the accuracy results of each scenario.
6 Conclusion
Reducing the dimensionality of the data improves the separation of samples into clusters.
However, dimensionality reduction has a limit in autoencoders, where the exaggerations
in reduction negatively affect the model prediction. It causes generalisation deficiencies.
Our approach remarkably reduced the dimension of the data to a single value. We
considered the reconstruction as a representative value. The clustering these values using
a fast algorithm like K-means instead of expensive computational algorithm improves the
performance. Our results show the proposed framework scored accuracy over 99%.
References
1. Ghorbani, A.A., Lu, W., Tavallaee, M.: Network Intrusion Detection and Prevention
Concepts and Techniques. Springer, New York (2010)
2. Mudzingwa, D., Agrawal, R.: A study of methodologies used in intrusion detection and
prevention systems (IDPS). In: 2012 Proceedings of IEEE SoutheastCon, pp. 1–6, 15–18
March 2012
3. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection for discrete sequences: a survey.
IEEE Trans. Knowl. Data Eng. 24(5), 823–839 (2012)
4. Hinton, G.E., Osindero, S., The, Y.W.: A fast learning algorithm for deep belief nets. Neural
Comput. 18, 1527–1554 (2006). https://doi.org/10.1162/neco.2006.18.7.1527
5. Krizhevsky, A., Sutskever, I., Hinton, G.: ImageNet classification with deep convolutional
neural networks. In: NIPS (2012)
6. Schmidhuber, J.: Deep learning in neural networks: an overview. Neural Netw. 61, 85–117
(2015)
7. Mohamed, A., Dahl, G., Hinton, G.: Acoustic modelling using deep belief networks. IEEE
Trans. Audio Speech Lang. Process. 20(1), 14–22 (2012)
8. Dong, B., Wang, X.: Comparison deep learning method to traditional methods using for
network intrusion detection. In: 2016 8th IEEE International Conference on Communication
Software and Networks (ICCSN), Beijing, pp. 581–585 (2016)
9. Sakurada, M., Yairi, T.: Anomaly detection using autoencoders with nonlinear dimension-
ality reduction. In: Rahman, A., Deng, J., Li, J. (eds.) Proceedings of the MLSDA 2014 2nd
Workshop on Machine Learning for Sensory Data Analysis (MLSDA 2014), p. 4, 8. ACM,
New York (2014). http://dx.doi.org/10.1145/2689746.2689747
10. Fiore, U., Palmieri, F., Castiglione, A., De Santis, A.: Network anomaly detection with the
restricted Boltzmann machine. Neurocomputing 122, 13–23 (2013)
11. Salama, M.A., Eid, H.F., Ramadan, R.A., Darwish, A., Hassanien, A.E.: Hybrid intelligent
intrusion detection scheme. In: Soft Computing in Industrial Applications, pp. 293–303.
Springer, Heidelberg (2011)
12. Zhai, S., Cheng, Y., Lu, W., Zhang, Z.: Deep structured energy based models for anomaly
detection. In: Maria Balcan, F., Weinberger, K.Q. (eds.) Proceedings of the 33rd
International Conference on Machine Learning - Volume 48 (ICML 2016), vol. 48,
pp. 1100–1109 (2016). JMLR.org
13. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.: A detailed analysis of the KDD CUP 99
data set. In: IEEE Symposium on Computational Intelligence for Security and Defense
Applications, CISDA 2009, pp. 1–6 (2009). https://doi.org/10.1109/CISDA.2009.5356528
Blockchain: A Research Framework for Data
Security and Privacy
current and future studies, and is titled ‘Themes of Data Security and Privacy through
Blockchain’. This part of the review will focus on synthesising various existing
blockchain studies that directly relate to the use of blockchain technology for the
maintaining data security and privacy. The Literature Review section is followed by the
last section that provides concluding remarks along with highlight of the authors’ future
work.
2 Literature Review
As stated before, this study identifies various themes and dimensions that collectively
define and scope current and future research in data security through blockchain.
Initially, the proposed research framework is based on the existing body of literature in
the fields, and the themes that the front-runner researchers have implied for their
research. For the current study, we have explored a set of initial themes as a starting
point. As one of the authors’ future work, the proposed research framework of the
current study will be updated cumulatively, as additional works are found in the
literature; hence the concept dynamic meta thematic model for blockchain research. To
achieve the above goal, we explored the relevant current knowledge landscape in order
to search for existing research works by early researchers that either directly or indi-
rectly relate to the topic of data security through blockchain. The current study partly
achieves this goal. More specifically, in an attempt to construct our proposed research
framework, we conduct a two-round review of the literature. In the first round, that is
the current study; we develop an initial categorisation scheme by identifying possible
themes or categories that define the current body of the literature in the field. In the
second round of review, in our future study, each existing study will be assessed by the
proposed framework to either be matched with an existing category/theme, or create a
new category/them, or update an existing one. The authors acknowledge the dynamic
nature of the proposed research framework, and for this reason, they intend to use the
current study as a basis for attracting collaborating researchers and future research
grants for maintaining the currency of the proposed research framework for all aspects
of Blockchain research.
The research methodology adopted for the current study is a conceptual systematic
review where the search criteria is a specific concept, and the goal of the review is to
develop a conceptual framework to guide empirical data collection and analysis [10].
To apply this methodology for the current study, we have collected relevant research
articles and major industry reports, and grouped them according to a set of pre-defined
concepts called categories or themes. In the current study, the primary theme at the
highest level is blockchain-based security, that has been used in the first round of the
review. During this first round of the review, we identified two sets of relevant research
categories: (i) blockchain technological and application perspectives, and (ii) specific
themes related to the data security and privacy through blockchain. These two are
discussed below:
Blockchain: A Research Framework for Data Security and Privacy 969
to some rules. Moreover, it is responsible for ordering the transaction. The whole
process could be divided into three phases: pre-prepared, prepared, and commit. In
each phase, a node would enter next phase if it has received votes from over 2/3 of all
nodes. So, PBFT requires that every node is known to the network.
Stellar Consensus Protocol (SCP) [25] is also a Byzantine agreement protocol.
There is no hashing procedure in PBFT. In PBFT, each node has to query other nodes
while SCP gives participants the right to choose which set of other participants to
believe. Based on PBFT, Antshares [26] has implemented their delegated byzantine
fault tolerance (dBFT). In dBFT, some professional nodes are voted to record the
transactions instead of all nodes.
Delegated proof of stake (DPOS) is a method similar to POS. In this algorithm,
miners get their priority to generate the blocks according to their stake. The significant
difference between POS and DPOS is that POS is a direct democratic while DPOS is a
representative democracy. Stakeholders elect their delegates to generate and validate a
block. With significantly fewer nodes to validate the block, the block could be con-
firmed quickly, making the transactions confirmed quickly. Meanwhile, the parameters
of the network such as block size and block intervals could be tuned. Additionally,
users do not need to worry about the dishonest delegates because the delegates could be
voted out quickly. Ripple is a consensus algorithm that utilises collectively-trusted
subnetworks within the more extensive network [27]. In the network, nodes are divided
into two types: a server for participating consensus process and client for only trans-
ferring funds. In contrast to that PBFT nodes have to ask every node in the network,
each Ripple server has a Unique Node List (UNL) to query.
In the next phase of our study, we plan to use the proposed research framework of
the current study and periodically reassess all upcoming works in the field of data
security through blockchain, and as a result, to enhance the proposed framework. Such
enhancement is defined by a need to either updating some of the existing themes or by
adding new themes to the research landscape. These iterative enhancements will
continue until the point where the last group of studies do not bring a major change to
the research framework. At that stage, one can claim to settle down of the emerging
blockchain technology about the security issues at academic levels.
References
1. Puthal, D., Malik, N., Mohanty, S.P., Kougianos, E., Yang, C.: The blockchain as a
decentralized security framework. IEEE Consum. Electron. Mag. 7, 18–21 (2018)
2. Goldfeder, S., Kalodner, H., Reisman, D., Narayanan, A.: When the cookie meets the
blockchain: privacy risks of web payments via cryptocurrencies. Proc. Priv. Enhanc.
Technol. 2018, 179–199 (2018)
3. Zheng, Z., Xie, S., Dai, H.-N., Chen, X., Wang, H.: Blockchain challenges and
opportunities: a survey. Int. J. Web Grid Serv. 14, 352–375 (2018)
4. Kosba, A., Miller, A., Shi, E., Wen, Z., Papamanthou, C.: Hawk: the blockchain model of
cryptography and privacy-preserving smart contracts. In: 2016 IEEE Symposium on Security
and Privacy (SP), pp. 839–858 (2016)
5. Akins, B.W., Chapman, J.L., Gordon, J.M.: A whole new world: income tax considerations
of the Bitcoin economy. Pitt. Tax Rev. 12, 25 (2014)
6. Zhang, Y., Wen, J.: An IoT electric business model based on the protocol of bitcoin. In:
2015 18th International Conference on Intelligence in Next Generation Networks (ICIN),
pp. 184–191 (2015)
7. Sharples, M., Domingue, J.: The blockchain and kudos: a distributed system for educational
record, reputation and reward. In: European Conference on Technology Enhanced Learning,
pp. 490–496 (2016)
8. Noyes, C.: Bitav: fast anti-malware by distributed blockchain consensus and feedforward
scanning. arXiv preprint arXiv:1601.01405 (2016)
9. Standards Australia: Roadmap for blockchain standards report (2017)
10. Seuring, S., Müller, M.: From a literature review to a conceptual framework for sustainable
supply chain management. J. Clean. Prod. 16, 1699–1710 (2008)
11. Park, J.H., Park, J.H.: Blockchain security in cloud computing: use cases, challenges, and
solutions. Symmetry 9, 164 (2017)
12. Axon, L.: Privacy-awareness in blockchain-based PKI (2015)
13. Zyskind, G., Nathan, O.: Decentralizing privacy: using blockchain to protect personal data.
In: 2015 IEEE Security and Privacy Workshops (SPW), pp. 180-184 (2015)
14. Chuen, D.L.K.: Handbook of Digital Currency: Bitcoin, Innovation, Financial Instruments,
and Big Data. Academic Press, Cambridge (2015)
15. Zheng, Z., Xie, S., Dai, H., Chen, X., Wang, H.: An overview of blockchain technology:
architecture, consensus, and future trends. In: 2017 IEEE International Congress on Big Data
(BigData Congress), pp. 557-564 (2017)
16. Dorri, A., Steger, M., Kanhere, S.S., Jurdak, R.: Blockchain: a distributed solution to
automotive security and privacy. IEEE Commun. Mag. 55, 119–125 (2017)
17. De Filippi, P.: The interplay between decentralization and privacy: the case of blockchain
technologies (2016)
974 F. Daneshgar et al.
18. Sarwar, M.U., Hanif, M.K., Talib, R., Sarwar, B., Hussain, W.: Data provenance for cloud
computing using watermark. Int. J. Adv. Comput. Sci. Appl. 8, 407–411 (2017)
19. Liang, X., Shetty, S., Tosh, D., Kamhoua, C., Kwiat, K., Njilla, L.: Provchain: a blockchain-
based data provenance architecture in cloud environment with enhanced privacy and
availability. In: Proceedings of the 17th IEEE/ACM International Symposium on Cluster,
Cloud and Grid Computing, pp. 468–477 (2017)
20. Dorri, A., Kanhere, S.S., Jurdak, R., Gauravaram, P.: Blockchain for IoT security and
privacy: the case study of a smart home. In: 2017 IEEE International Conference on
Pervasive Computing and Communications Workshops (PerCom Workshops), pp. 618–623
(2017)
21. Huckle, S., Bhattacharya, R., White, M., Beloff, N.: Internet of things, blockchain and shared
economy applications. Proc. Comput. Sci. 98, 461–466 (2016)
22. Castro, M., Liskov, B.: Practical Byzantine fault tolerance. In: OSDI, pp. 173–186 (1999)
23. Nakamoto, S.: The proof-of-work chain is a solution to the Byzantine generals’ problem.
The Cryptography Mailing List, November 2008
24. Cachin, C.: Architecture of the hyperledger blockchain fabric. In: Workshop on Distributed
Cryptocurrencies and Consensus Ledgers (2016)
25. Mazieres, D.: The stellar consensus protocol: a federated model for internet-level consensus.
Stellar Development Foundation (2015)
26. Zhu, H., Zhou, Z.Z.: Analysis and outlook of applications of blockchain technology to
equity crowdfunding in China. Financ. Innov. 2, 29 (2016)
27. Schwartz, D., Youngs, N., Britto, A.: The ripple protocol consensus algorithm. Ripple Labs
Inc White Paper (2014)
Environmental Monitoring Intelligent System
Using Wireless Nanosensor Networks
1 Introduction
2 PlantMONT
The expected capabilities and peculiarities of NSNs and the terahertz channel require
the development of novel, reliable and yet energy efficient medium access and com-
munication protocols. In this project, new medium access control (MAC), addressing,
network discovery and association and routing protocols will be designed, imple-
mented and investigated. PlantMONT has the following concrete objectives:
Integer Linear Program (MILP) model to solve it. There are few attempts in the
literature to address multi-access and channel sharing problem in NSNs environment,
however most of the research do not take into consideration the unique characteristics
of such networks and thus, need to be improved further. Jornet et al. suggested the first
MAC protocol for WNSN, called PHLAME [15]. The proposed framework design is
based on a new modulation schema called RD TS-OOK, an extended version of TS-
OOK [16]. PHLAME uses the advantages of the low-weight channel coding protocol
proposed in [17]. By reducing the code weight, that is, encoding the information using
more logical 0’s than logical 1’s, both molecular absorption noise and interference can
be mitigated without affecting the throughput of the WNSNs. PHLAME provides joint
selection of the optimal communication parameters and channel coding scheme for
both the sender and the receiver. However, implementation feasibility and energy
efficiency evaluation of the method are still open questions: First, the network scala-
bility might be an issue for PHLAME, as performance of the proposed approach has
only been investigated for a maximum of 10 nanomotes per square millimeter, which
might not be a realistic scenario for some applications where thousands of nanosensors
per square millimetre are needed. Second, PHLAME relies on the low-weight channel
coding, which might not be applicable for WNSNs in some environments where the
codeword is hard to define for the transmitted data due to natural randomness in the
environment. Finally, the handshake process in the PHLAME may exhaust the energy
of nanosensors, limiting the throughput of the network, and nanosensors might not
have enough computational resources to dynamically find the optimal communication
parameters. To alleviate the last restriction of the PHLAME, Wang et al. extended the
original PHLAME to increase the throughput and lifetime of WNSNs by jointly
optimizing the energy harvesting and consumption processes [18]. In other research,
Piro et al. attempted to adopt the existing MAC protocols, mainly ALOHA and carrier
sense multiple access (CSMA), for WNSNs and investigate their performances [19].
2.3 Routing
Nanosensors may have to answer a specific query from a command center or may need
to report new events in a push-based fashion. This flow of information (in two-
directions) requires the establishment of routes. Due to the very limited transmission
range (in the order of few tens of millimeters) of nanosensors, multi-hop communication
will be the standard way to communicate between senders and receivers. It is worth to
note that the creation and the management of multi-hop paths from each nanosensor to
its reference nanorouter (cluster head or macrosink) is a very challenging open issue,
which could be very critical in those scenarios entailing some degree of mobility. Tiny
nanosensor devices randomly deployed in an open field can be easily taken away by an
air flow. This may be seen as a shortcoming but can also be exploited in a beneficial
way. All main aspects related to the communications among nanosensors, the interac-
tions between them and nanorouters, and the interface/gateway between a WNSN and
the Internet (command center) will be deeply studied in this project to discover the most
promising and feasible solutions. We will investigate novel Inter- and Intra-network
(plant) routing mechanisms that consider the particular energy requirements and
mobility pattern of nansensors. Also, End-to-end reliability in NSNs has to be guar-
anteed both for the packets going from a remote command center to the nanosensors, as
well as for the packets coming from the nanosensors to a common nanosink (nanor-
outer). We will investigate multi-path routing, selecting the optimal path, forming
clusters, nanorouter/sink optimal placement, and data dissemination and aggregation
techniques. We will design a multi-objective function to address several routing metrics
such as energy consumption, end-to-end delay, throughput, and communication relia-
bility. Pierobon et al. proposed the first routing protocol for WNSNs to optimize the use
of harvested energy to guarantee continuous operation of WNSNs while improving the
overall network throughput [20]. Due to the resource restriction of the nanosensors, in
this work, a hierarchical cluster-based structure is used in which the WNSN is formed
from a few clusters, and each cluster is coordinated by a nanoscale controller, a nano
device with more advanced capabilities than a nanosensor. Later, Liaskos et al. proposed
a metaheuristic-based, selective flooding dissemination scheme for nano-networks [21],
which was later refined in terms of complexity [22]. Tsioliaridou et. al. proposed a joint
peer-to-peer routing and geo-addressing approach, CORONA, for NSNs [23]. COR-
ONA is a stateless and scalable nano-routing approach. It proposes a packet flood
approach, contained within the area defined by the minimal and maximal anchor dis-
tances of the communicating node-pair. However, routing protocol design for nano-
networks is in its very early stages and lots of effort needed here to address the different
routing issues. As stated above, nanosensors are not commercially available yet, and
980 N. Ababneh et al.
thus, real experimentations on nanosensor devices are not possible. Moreover, since
WNSNs are too complex for traditional analytical methods to provide an accurate
modelling and understanding of system behavior, network simulators (i.e., ns-2 and
Nano-Sim based on ns-3) and Mathematical Programming-based optimization tech-
niques such as ILP and constraint programming will be used in this project. The ns-2 and
ns-3 simulators are the most prominent tools to develop network, protocol and topology;
they provide substantial supports for designing physical layer, MAC, and Routing
protocols, and investigating how various protocols perform with different configurations
and topologies in a controlled manner. To the best of our knowledge, these two
methodologies/approaches are never been used in the literature so far.
3 Implementation
Three Work Packages (WP) are necessary to achieve the research goals of PlantMONT
(detailed in Sect. 2), and dedicate one other to dissemination and exploitation (and
public engagement) of results:
WP1: This work package deals with the development of the models (Objective 2.1).
WP2: This work package deals with the development of models and solutions for
objectives 2.2 and partially 2.3.
WP3: Having a good understanding of MAC, intra-network routing/dissemination,
and network topology/deployment, we are ready to look at (Objective 2.3):
4 Conclusion
References
1. Akyildiz, I.F., Miquel Jornet, J.: Electromagnetic wireless nanosensor networks. Nano
Commun. Netw. 1(1), 3–19 (2010)
2. Ababneh, N., Viglas, A., Labiod, H., Boukhatem, N.: ECTC: energy efficient topology
control algorithm for wireless sensor networks. In: Proceedings of the IEEE International
Symposium on a World of Wireless Mobile and Multimedia Networks (WoWMoM 2009),
Kos Island, Greece, June 2009
Environmental Monitoring Intelligent System 981
3. Ababneh, N.: Radio irregularity problem in wireless sensor networks: a new experimental
results. In: Proceedings of the 32nd IEEE Sarnoff Symposium (Sarnoff 2009), Princeton,
USA, April 2009
4. Ababneh, N., Ortiz, A.M., Timmons, N., Morrison, J.: Energy-efficient multicast tree
construction protocol for real-time data streaming in WSNs. In: Proceedings of the IEEE
International Conference on Wireless and Mobile Computing, Networking and Communi-
cations (WiMob 2012), Barcelona, Spain, 8–10 October 2012
5. Ababneh, N., Timmons, N., Morrison, J.: An adaptive bandwidth allocation scheme for data
streaming over body area networks. In: Proceedings of WWIC (2012)
6. Ababneh, N., Timmons, N., Morrison, J.: Cross-layer optimization protocol for guaranteed
data streaming over wireless body area networks. In: Proceedings of the International
Wireless Communications and Mobile Computing Conference (IWCMC 2012), Cyprus, 27–
31 August 2012
7. Yonzon, C.R., Stuart, D.A., Zhang, X., McFarland, A.D., Haynes, C.L., Van Duyne, R.P.:
Towards advanced chemical and biological nanosensors—an overview. Talanta 67(3), 438–
448 (2005). Nanoscience and Nanotechnology
8. Li, C., Thostenson, E.T., Chou, T.W.: Sensors and actuators based on carbon nanotubes and
their composites: a review. Compos. Sci. Technol. 68(6), 1227–1249 (2008)
9. Cui, Y., Wei, Q., Park, H., Lieber, C.M.: Nanowire nanosensors for highly sensitive and
selective detection of biological and chemical species. Science 293(5533), 1289–1292
(2001)
10. Jornet, J., Akyildiz, I.F.: Graphene-based nano-antennas for electromagnetic nanocommu-
nications in the terahertz band. In: European Conference on Antennas and Propagation
(EuCAP), pp. 1–5 (2010)
11. Heil, M., Bueno, J.C.S.: Within-plant signaling by volatiles leads to induction and priming
of an indirect plant defense in nature. Proc. Natl. Acad. Sci. U.S.A. 104, 5467–5472 (2007)
12. Heil, M., Ton, J.: Long-distance signalling in plant defence. Trends Plant Sci. 13(6), 264–
272 (2008)
13. Afsharinejad, A., Davy, A., Jennings, B.: Dynamic channel allocation in electromagnetic
nanonetworks for high resolution monitoring of plants. Nano Commun. Netw. 7, 2–16
(2016)
14. Akyildiz, I.F., Jornet, J.M.: The Internet of nano-things. IEEE Wirel. Commun. 17(6), 58–63
(2010)
15. Jornet, J.M., Pujol, J.C., Pareta, J.S.: PHLAME: a physical layer aware MAC protocol for
electromagnetic nanonetworks in the Terahertz Band. Nano Commun. Netw. 3(1), 74–81
(2012)
16. Jornet, J.M., Akyildiz, I.F.: Information capacity of pulse-based Wireless Nanosensor
Networks. In: 2011 8th Annual IEEE Communications Society Conference on Sensor, Mesh
and Ad Hoc Communications and Networks, pp. 80–88 (2011)
17. Jornet, J.M., Akyildiz, I.F.: Low-weight channel coding for interference mitigation in
electromagnetic nanonetworks in the Terahertz Band. In: 2011 IEEE International
Conference on Communications (ICC), pp. 1–6 (2011)
18. Wang, P., Jornet, J., Malik, A., Akkari, N., Akyildiz, I.F.: Energy and spectrum-aware MAC
protocol for perpetual wireless nanosensor networks in the Terahertz Band. Ad Hoc Netw.
11(8), 2541–2555 (2013)
19. Piro, G., Abadal, S., Mestres, A., Alarcón, E., Solé-Pareta, J., Grieco, L.A., Boggia, G.:
Initial MAC exploration for graphene-enabled wireless networks-on-chip. In: Proceedings of
ACM NANOCOM, Atlanta, GA (2014)
982 N. Ababneh et al.
20. Pierobon, M., Jornet, J., Akkari, N., Almasri, S., Akyildiz, I.F.: A routing framework for
energy harvesting wireless nanosensor networks in the Terahertz Band. Wirel. Netw. 20,
1169–1183 (2013)
21. Liaskos, C., Tsioliarido, A.: A promise of realizable, ultra-scalable communications at nano-
scale: a multi-modal nano-machine architecture. IEEE Trans. Comput. 64, 1282–1295
(2015)
22. Tsioliaridou, A., Liaskos, C., Ioannidis, S., Pitsillides, A.: Lightweight, self-tuning data
dissemination for dense nanonetworks. Nano Commun. Netw. 8, 2–15 (2015)
23. Tsioliaridou, A., Liaskos, C., Ioannidis, S., Pitsillides, A.: CORONA: a coordinate and
routing system for nanonetworks. In: NANOCOM (2015)
The 1st International Workshop on
Multi-clouds and Mobile Edge
Computing (M2EC-2019)
Situation Detection on the Edge
1 Introduction
Mobile Edge Computing (MEC) enables a computing and storage infrastructure pro-
visioned closely to the end-users at the edge of a cellular network. Combining MEC in
multi-cloud infrastructures can help to combat latency challenges imposed by cloud-
centric architectures. However, edge devices are highly dynamic in nature: they are not
as reliable as server and cloud computing resources; they computing capacity is limited
and varies greatly depending on their workload; their operating environment (tem-
perature, humidity, etc.) may impact their performance; they often include sensors
which send data at a very high rate which can sometimes swamp the available network
bandwidth. Perception of these elements in the environment of edge devices within a
volume of time and space and the comprehension of their meaning is typically referred
to as ‘situation’ [6, 7]. Situation detection can enhance the capacity to manage edge
resources effectively as part of a computing environment.
Situations in edge computing infrastructures are highly related to the current status
and context of edge devices and behavior of deployed applications. Situations are in a
rich structural and temporal relationship, and they are dynamic by definition, continu-
ously evolving and adapting. To cope with the dynamicity of situations, one needs to
sense and process data in large volumes, in different modalities [18]. To realize systems
for Situation Awareness (SA) “individual pieces of raw information (e.g. sensor data)
should be interpreted into a higher, domain-relevant concept called situation, which is an
abstract state of affairs interesting to specific applications” [13]. The power of using
‘situations’ lies in their ability to provide a simple, human-understandable representation
of, for instance, sensor data [13]. In the context of dynamic computing systems, a
situation is defined as an event occurrence that might require a reaction [1]. In com-
puting, SA is the capability of the entities of the computing environment to be aware of
situation changes and automatically adapt themselves to such changes to satisfy user
requirements, including security and privacy [17]. SA is one of the most fundamental
features to support dynamic adaptation of entities in pervasive computing environments.
The research objective of our work has been to design and develop a Situation
Detection Mechanism (SDM) capable of processing in real-time events generated by
the edge infrastructure and detecting situations where the edge infrastructure requires
an adaptation. Moreover, the SDM should be modular and easily deployable on
heterogeneous edge infrastructures. The deployment flexibility of SDM is quite
important since it allows using it on as many as possible different types of edge devices.
SDM is part of the PrEstoCloud dynamic and distributed software architecture that
manages cloud and fog resources proactively, while reaching the extreme edge of the
network for an efficient real-time BigData processing (http://prestocloud-project.eu/).
The rest of the paper is organized as follows. Section 2 discusses enabling tech-
nologies and works related to situation detection and awareness for enterprise decision
making with an emphasis on supporting proactivity. Section 3 outlines the proposed
situation model and described the approach followed to realise situation detection.
Section 4 presents a scenario in which SDM has been evaluated. Section 5 discusses
the main findings of our work and our future plans.
2 Related Work
Unsupervised learning methods can help ameliorate this challenge but there are limited
works reported in the literature. Learning-based methods can cope well with uncertainty
when trained with noisy real-world data [18].
both the deployed big data intensive application and the underlying multi-cloud and
edge resources. Thus, all MonitoringInformation is based on the Processing, Host-
ingEdgeNodes, HostingCloudNodes and current Workload detected through the
appropriate software, hardware and workload related monitoring probes, respectively.
Both Processing and Workload are expressed based on the BigDataVocabulary in order
to abstractly map types of big-data streams to big data processing services types
revealing their importance for the detected PrEstoCloudSituations.
control and customize the execution environment. For example edge devices may have
very low computational resources or a very restricted (due to security reasons)
environment for custom applications. Very often those devices have low network
bandwidth, unpredictable disconnections from the network and data transmission spikes
that are caused by external events (such as social events, weather conditions or other). In
this environment we need infrastructure and mechanisms for data-driven event detec-
tion. Therefore, we opted for an approach that relies on Complex Event Processing
technologies, which are capable of processing in real-time a large number of events
generated by a variety of distributed cloud and edge computing resources as well as
other data generating sensors. A complex event is an event derived from a group of
events using either aggregation of derivation functions. Information enclosed in a set of
related events can be represented (i.e., summarized) through such a complex event.
The primary input for Situation Detection Mechanism (Fig. 2) is considered any
health status event or application performance-related event transmitted to the Com-
munication and Message Broker (RabbitMQ). Such events are exploited by SDM in
order to reveal problematic situations with respect to the state of the cloud and edge
resources used for hosting big data-intensive applications or to the performance state of
the application itself. Events from the cloud infrastructure (physical and virtual
machines, containers, applications, services, etc.) and edge devices (mobile phones,
routers, IoT devices) are published as events to the Broker in specific topics. One or
more SDM service instances subscribe to the desired topics and receive streams of
events that contain up-to-date information about the current state of those entities (e.g.
used RAM, CPU consumption, disk I/O, requests per second, etc.). SDM instances
process these events based on the supplied CEP rules which are defined in order to
detect interesting situations. Several SDM instances can be used in parallel or in series
in order to process the incoming event streams. High level situations can be detected by
processing low-level situations from many SDM instances.
4 Evaluation
We have evaluated SDM using two different CEP engines, Drools and Siddhi. The
software and hardware configuration is as follows: Hardware (KVM Virtual Machine
with 4 cores and 8 GB RAM running on a server with Intel Xeon E7 @ 2.4 GHz CPU),
Software (SDM services run under Ubuntu 17.10, Docker version 17.12.0-ce, build
c97c6d6, Docker-compose version 1.19.0, build 9e633ef, OpenJDK Runtime Envi-
ronment, Siddhi version v4.0.0 with RabbitMQ extension v1.0.14, Drools version
6.5.0. Final, RabbitMQ 3.7.5 (Docker image rabbitmq3.7.5-management). We used the
RabbitMQ load-testing tool to generate and publish events (https://github.com/
rabbitmq/rabbitmq-perf-test). We used only RabbitMQ and two SDM instances, one
implemented with the Drools CEP library and one implemented with the Siddhi library.
With PerfTest we can select the number of event producers, the length of the period
that we want to send events, the frequency with which the event producers should
generate events and the payload of the events (from a list of files). The AMQP
exchange name and the topic are also configurable. With a Java Management Exten-
sions (JMX) tool such as JConsole or VisualVM (https://visualvm.github.io/), we
monitored the Drools and Siddhi version of SDM.
We ran (with docker-compose) one Siddhi CEP engine and one Drools CEP engine
in parallel and configure them to subscribe to the same AMQP exchange and topic. In
this way both CEP engines received the same events from PerfTest. First we run
PerfTest for 60 s with increasing number of event producers that send one event per
second. The payload of the events is a JSON file that contains different values of two
attributes named “memory” and “cpu”, (without any timestamp for simplification
reasons). Both Drools and Siddhi were configured to produce every 10 s two events
containing: (i) the average CPU and MEMORY (during the last 10 s); (ii) the number
of MEMORY and CPU events that it received (during the last 10 s).
Situation Detection on the Edge 991
Fig. 3. SDM load testing with PerfTest. CPU utilization of Drools-based implementation vs
Siddhi-based implementation (500, 500, 1000 events/s)
We compared the CPU consumption of Drools and Siddhi while sending 500
events per second (twice) and 1000 events per second, for two consecutive periods of
60 s. We can see clearly in Fig. 3 that the Siddhi-based implementation of SDM has
much lower total CPU utilization than the Drools-based implementation which
increases in a bigger proportion as the rate of incoming events increases.
Fig. 4. SDM load testing with PerfTest. CPU utilization of Drools-based implementation vs
Siddhi-based implementation (1250 to 3000 events/s)
In Fig. 4, we continue the same experiment with increasing number of events per
second (generated by PerfTest): 1250, 1500, 1750, 2000, 2250, 3000. It is again clear
that Siddhi has much lower CPU utilisation. It is also notable that after 1500 events per
second the Drools-based implementation of SDM queues the incoming messages and
992 N. Papageorgiou et al.
Fig. 5. SDM load testing with PerfTest. Used memory of Drools-based implementation vs
Siddhi-based implementation (500, 500, 1000 events/s)
Further, we compared the expressivity of the Drools and Siddhi rule languages in
expressing situations. Siddhi rule language is SQL-like while Drools language follows
the paradigm of Event-Condition-Action (ECA) rules. Both rule languages have dif-
ferent but straightforward expressions for dealing with simple functions like the
average, the minimum or the maximum of a metric over a period of time. When we
want to aggregate over different objects and group the results over a group key,
according to our knowledge, Drools rule language is not as expressive as Siddhi rule
language. In the following examples we show how we can detect a situation where
more than 20 requests for the same URL are detected in a period of time of 10 s. In
Drools we have to produce an event (RequestAlert) which expires in 10 s and insert it
in the fact base in order to prevent Drools to continuously produce the situation after
the first time it was detected.
Situation Detection on the Edge 993
/* Drools rule to alert when more than N=20 requests for the same url
are detected in a period of 10 seconds.*/
declare RequestAlert
@role( event )
@timestamp( ts )
@expires ( 10s )
ts: Date @key
url: String @key
end
rule "request_url alert"
timer ( int: 10s 10s )
when
$e1 : SquidEvent( $url : request_url )
not RequestAlert( url == $url )
$c : Number(intValue > 20 ) from accumulate (
$e2 : SquidEvent( this != $e1, request_url == $url ) over
window:time(10s), count( $e2 ) )
then
insert ( new RequestAlert( new Date(), $url ) );
log( "[REQURL] " + $c + " requests to the same url in 10s" );
end
With Siddhi we can group by the attribute “rurl” (that corresponds to the request
URL) and check every 10 s (with #window.timeBatch) for each one if the total is over
20 (having cnt > 20).
/* Siddhi rule to alert when more than N=20 requests for the same url
are detected in a period of 10 seconds.*/
from squidStream#window.timeBatch(10 sec)
select count(rurl) as cnt , rurl
group by rurl
having cnt > 20
insert into rurlStream;
from rurlStream
select str:concat("[REQURL] " , cnt, " requests to the same url
in 10s" ) as msg
insert into msgStream;
5 Conclusions
SDM has been designed and developed focusing on providing detection capabilities for
situations pertaining to edge resources computing capabilities. In implementing SDM,
we followed a specification-based approach. We designed the SDM component so as it
is modular and can be easily deployed as a Docker container or a set of Docker
containers. Moreover, we designed SDM to be independent of CEP libraries and we
have shown that it can operate with both the Siddhi and Drools CEP libraries. The
deployment flexibility of SDM is quite important since it allows using the Complex
Event Processing engine of choice based on the processing capabilities required in each
case and the prior expertise regarding a certain Complex Event Processing engine.
994 N. Papageorgiou et al.
References
1. Adi, A., Etzion, O.: Amit - the situation manager. VLDB J. 13(2), 177–203 (2004)
2. Amazon (2018). https://aws.amazon.com/autoscaling
3. Chen, H., Finin, T., Joshi, A.: An ontology for context-aware pervasive computing
environments. Knowl. Eng. Rev. 18(3), 197–207 (2004)
4. Cohen, N.H., Lei, H., Castro, P., Davis II., J.S., Purakayastha, A.: Composing pervasive data
using iQL. In: Proceedings of the Fourth IEEE Workshop on Mobile Computing Systems
and Applications, WMCSA 2002, pp. 94–104 (2002)
5. Haghighi, P.D., Krishnaswamy, S., Zaslavsky, A., Gaber, M.M.: Reasoning about context in
uncertain pervasive computing environments. In: Proceedings of the 3rd European
Conference on Smart Sensing and Context, EuroSSC 2008, pp. 112–125. Springer,
Heidelberg (2008)
6. Endsley, M.: Designing for Situation Awareness: An Approach to User-Centered Design,
2nd edn. CRC Press, Boca Raton (2016)
7. Franke, U., Brynielsson, J.: Cyber situational awareness - a systematic review of the
literature. Comput. Secur. 46, 18–31 (2014)
8. Google Cloud (2018). https://cloud.google.com/compute/docs/autoscaler/
9. Gray, P.D., Salber, D.: Modelling and using sensed context information in the design of
interactive applications. In: Proceedings of the 8th IFIP International Conference on
Engineering for Human-Computer Interaction, EHCI 2001, pp. 317–336. Springer, London
(2001)
10. Gu, T., Chen, S., Tao, X., Lu, J.: A non supervised approach to activity recognition and
segmentation based on object-use fingerprints. Data Knowl. Eng. 69(6), 533–544 (2010)
11. Kubernetes (2018). https://kubernetes.io/docs/tasks/run-application/horizontal-pod-auto-
scale-walkthrough/
Situation Detection on the Edge 995
12. Lei, H., Sow, D.M., John, S., Davis, I., Banavar, G., Ebling, M.R.: The design and applications
of a context service. SIGMOBILE Mob. Comput. Commun. Rev. 6(4), 45–55 (2002)
13. Loia, V., D’Aniello, G., Gaeta, A., Orciuoli, F.: Enforcing situation awareness with granular
computing: a systematic overview and new perspectives. Granular Comput. 1(2), 127–143
(2016)
14. Loke, S.W.: Incremental awareness and compositionality: a design philosophy for context-
aware pervasive systems. Pervasive Mob. Comput. 6(2), 239–253 (2010)
15. OpenStack (2018). https://docs.openstack.org/senlin/latest/scenarios/autoscaling_heat.html
16. Ranganathan, A., Al-Muhtadi, J., Campbell, R.H.: Reasoning about uncertain contexts in
pervasive computing environments. IEEE Pervasive Comput. 03(2), 62–70 (2004)
17. Yau, S.S., Liu, J.: Hierarchical situation modeling and reasoning for pervasive computing.
In: The Fourth IEEE Workshop on Software Technologies for Future Embedded and
Ubiquitous Systems, SEUS 2006/WCCIA 2006, pp. 6-pp, April 2006
18. Ye, J., Dobson, S., McKeever, S.: Situation identification techniques in pervasive
computing: a review. Pervasive Mob. Comput. 8(1), 36–66 (2012)
A Context-Aware Service for Authorizing
Multi-cloud Deployments
1 Introduction
In recent times, there is an abundance of cloud services offered from a vastly increasing
number of cloud providers, each one with its own advantages or shortcomings. Finding
opportunities where the benefits of each provider are exploited at the same time, while
coping with the ever-increasing requirements of the modern data-intensive cloud
applications, has become imperative [1]. A generic challenge is to overcome scala-
bility, resiliency, and security issues faced by big data and data-intensive applications
on distributed platforms by using transparent and optimized multi-cloud resource
provisioning [1]. One of the critical issues towards tackling this generic challenge is to
design and develop the appropriate methods and tools for adequately protecting the
platform components that make possible the optimized application placement on multi-
clouds. We use as an example the platform presented in [2] where decision and
optimization components (called Upperware) are used for ingesting user-defined
application placement requirements, constraints and optimization goals in order to
provide optimized decisions on initial placement or reconfigurations of data-intensive
applications on multi-clouds. A second part of such a platform involves a set of
components that interface with the required providers for commissioning virtual
resources and deploying application components according to the initial placement or
the reconfiguration decisions of the Upperware components. This platform [2] follows
a model-driven engineering approach, where models as first class citizens, initially
capture placement requirements and optimization goals, providing valuable input for
analysis and decision making with respect to multi-cloud application placement and
last they are transformed to executable service graphs that guide the placement or
reconfiguration actions across clouds.
Nevertheless, the list of the main risks associated with clouds is very long, and even
longer for multi-clouds [3]. At the top of this list, we find the improper identity
management, insufficient credentials and access control, unsecured application pro-
gramming interfaces (APIs) [4]. Although there are several security-related topics that a
platform that undertakes applications placement on multi-clouds should cover (e.g.
cloud providers credentials management, component authentication, encrypted com-
munication etc.), in this work we are focusing on an advanced authorization module,
based on the Attribute-based Access Control (ABAC) paradigm. This mechanism is
designed to provide access control and pre-authorization of any deployment plans that
may be implemented by such a platform. We note that such security related capabilities
are applicable to any application placement scenario but we mainly focus on data-
intensive applications due to the inherent additional requirements for efficiency that
introduce. The introduction of this kind of authorization capabilities, ensures that only
eligible entities (users or components) can access protected platform resources and
apply certain operations on them. Each access attempt to a resource is checked against
a set of access control policies. During authorization checking, various stated and
contextual information must be considered; this information relates to the requestor (i.e.
user or component), the resource being accessed (e.g. data, methods etc.), the attempted
operation as well as other environment data (e.g. date, component operational status,
etc.). Beyond access control, the authorization infrastructure should be consulted on
whether a given application deployment plan, generated by the appropriate decision
and optimization components, complies with a set of deployment policies. Such
policies may encompass constraints and limitations referring to application deployment
(e.g. total cost or number of virtual machines deployed or location of resources). In this
way, the described security enhancement of the platform leads to a protected operation
of all critical software that makes decisions and implements placement or reconfigu-
ration actions regarding multi-cloud deployments.
The rest the paper involves a relevant state-of-the-art analysis which is provided in
Sect. 2, while in Sect. 3, we discuss the requirements of the developed service called
Melodic Authorization Service. In Sect. 4, we provide the design details of this
mechanism, while in Sect. 5 the implementation specifics of this service are described.
Last, we conclude with Sect. 6, where the next steps of this work are discussed.
Several access control models have been proposed in the literature and used in software
products. These models provide a framework and a method of how resources,
requestors, operations, and rules may be combined to produce and enforce an access
control decision. Some of the most well-known models are the Discretionary Access
998 Y. Verginadis et al.
Control (DAC) [5], the Mandatory Access Control (MAC) [6], the Identity-Based
Access Control (IBAC) [6], the Role-Based Access Control (RBAC) [7] and the Attri-
bute-Based Access Control (ABAC) [7]. Out of the most prominent paradigms is the
ABAC, especially if we consider the dynamic characteristics of the cloud computing
domain. ABAC uses policies that comprise rules, which in turn comprise logical
conditions on several different attributes. Typically, each rule contains at least a con-
dition (i.e. boolean expression) and a decision (permit or deny) to reach when the
condition is true. Policies combine the outcomes of rules and yield the final decision
using certain outcome combination methods. Attributes can be properties of the
requesting entity, of the resource being accessed, of the operation requested, or any
other contextual information found in the so-called external environment enclosing the
interacting entities. Thus, these attributes are not limited in number or in type as in
previous paradigms (e.g. RBAC which only considers the role of the access requesting
entity).
There are a few reference implementations of the ABAC model, but among the
most important ones are the eXtensible Access Control Markup Language (XACML)
and the Next Generation Access Control (NGAC) [8]. It is important to note that
A Context-Aware Service for Authorizing Multi-cloud Deployments 999
XACML seems to be the most widely used ABAC implementation model, since it
enjoys worldwide industrial adoption in sectors like banking, healthcare, and insurance
[9]. XACML is an XML-based, open-standard language promoted by OASIS, for
expressing authorization policies and querying access to resources. Evaluating an
access request to a resource, with regards to an XACML policy, may result in one of
these four values: Permit, Deny, Indeterminate (an error occurred or needed values
were missing) or Not Applicable (no related policy found). The XACML specification
defines five main components that handle access decisions; the Policy Enforcement
Point (PEP), Policy Administration Point (PAP), Policy Decision Point (PDP), Policy
Information Point (PIP), and a Context Handler (CH) [8]. Since XACML introduction,
several compliant frameworks have been developed and offered, both as open source
software (Table 1) as well as proprietary products.
We note that we have thoroughly discussed the numerous limitations of previous
works that do not follow the ABAC paradigm or its reference implementations for
coping with the challenges of data authorization in dynamic multi-cloud environments
[4]. In addition, this work proceeds further than previous efforts that try to overcome
the basic syntactic descriptions by capturing the knowledge that lurks behind policies
and rules in the sense that the relations between the attributes used for access control
are considered [4]. To the best of our knowledge such a context-aware access control
mechanism has not been implemented for securing the model-driven engineering
interactions, within platform components that enable multi-cloud deployments and
reconfigurations. Such a mechanism aspires to alleviate the repercussions of any
potential components’ compromise, coming from external adversaries, by authorizing
in a context-aware manner any updates on the models used.
can be of various types and originate from diverse sources. The Metadata Schema
introduced in [1], provides a classification of these information types, in its Context-
aware Security model (e.g. permission types, physical or network location of an
incoming request, time-related patterns based on normal access requests, etc.). This
classification can act as a common vocabulary for collecting and leveraging informa-
tion for authorization purposes.
The second objective relates to the enforcement of policies and limitations
regarding the deployment of cloud applications and their data across various cloud
providers. By design, a decision-making component in Upperware is responsible for
producing correct deployment plans conforming to any given constraints and limita-
tions. However, a compromised component, due to a cyber-attack, could possibly yield
invalid deployment plans. Therefore, a precautionary validation step, before the actual
deployment, would mitigate the risk of deploying an application in a non-conformant
manner. We refer to this step as pre-authorization. The enforced limitations can be
regulatory (e.g. data not allowed to leave the EU), corporate, as well as budget-,
resource- or security-related. A pre-authorization policy could for example, pose a limit
on the number of virtual machines deployed on a cloud provider, or require the storage
of data of a certain type to be stored in nodes located only in the EU. These two
objectives are quite different in their business purpose and involve different autho-
rization rules. However, the same authorization capabilities and toolset can be used in
order to achieve both of them.
In this section the details of the developed authorization service approach are high-
lighted. Specifically, we discuss the conceptual architecture of the system and provide
details on how the contextual attributes are managed by this system.
It is important to note that we have adopted the ABAC model for this service,
according also to the analysis provided in Sect. 2. Specifically, based on the infor-
mation provided in Table 1, we opted to use WSO2 Balana engine for XACML 3.0
implementation. However, replacing it with another alternative is expected to be a
relatively straightforward task, since the XACML policy engine resides inside the PDP
component of the XACML architecture. Furthermore, the pluggable design of the
server will allow the easy replacement of plugins pertaining to the specific policy
engine with new ones.
relaying the extracted information. If the server returns a positive decision, the access
request processing flow resumes. Otherwise, the access is prevented. PEP is provided
as an authorization service client library, which is embedded in the platform compo-
nents that should be protected. The communication to PDP is achieved using the
REST API exposed by the Authorization Server, over an encrypted transport layer
security (TLS) connection.
These PEPs communicate through an appropriate load balancing to a certain
Authorization Server. Depending on the use case, multiple Authorization Servers can
be used for scaling the access control capabilities that are provided by our system.
Specifically, an Authorization Server comprises a PDP, a Context Handler and several
PIPs. Each PDP constitutes a web service that provides a RESTful API for receiving
access request information from PEP’s, evaluating them against policies and eventually
authorizing or declining the certain access request. For this purpose, PDP contains a
policy evaluation engine, namely the WSO2 Balana. Upon configuration, PDP will first
invoke the Context Handler to collect additional (contextual) information from the
request or the environment, and then evaluates the incoming request against a number
of pre-defined access control policies. Several PDP nodes may coexist in a cluster to
achieve high availability, fault tolerance and fast response times. Typically, all PDP
nodes share the same configuration and the same policy repository. A second com-
ponent embedded in the Authorization Server is the Context Handler. Upon activation,
it invokes the configured plugins to collect additional information (as attributes) about
the context of the request. This contextual information is subsequently stored in a PIP
(in order to become available during the policy evaluation and access control decision.
Furthermore, the Context Handler receives platform or environment-related context
from Context Collectors. The PIP is responsible for providing values for all the
involved attributes that participate in the access control rules and policies, based on
which an incoming request is permitted or denied. Specifically, the PDP’s policy
evaluation engine, while processing a request, might require attribute values not con-
tained in the request itself. In such case, it invokes PIP plugins to retrieve the required
attributes that constitute key-value pairs, where keys are typically in the form of
Uniform Resource Names (URN).
The PAP is implemented in order to manage and store the XACML policies. These
policies are stored in a shared place accessible by all PDPs. The architecture is com-
plemented by a number of Context Collectors (CC), i.e. independent and domain-
specific application components that aim at continuously collecting information about
the multi-cloud management platform and its environment. Thus, the CCs are essen-
tially mechanisms for collecting contextual information from the environment and
propagate it to the Context Handlers. It is expected that different context information,
and thus context collectors, will be needed in different multi-cloud deployments based
on the application domain. Last, appropriate load balancing capabilities have been
fused into our system. Specifically, the PDP Load-Balancing stands between PEP
clients and the PDP nodes to prevent flooding of one or more of the Authorization
Server instances and allowing the proper scaling of the system. This is implemented at
the PEP-side as a dynamically configured list of PDP endpoints that are contacted
either in successive order (round-robin) or selected randomly. Similarly, Context
Handler load balancing capabilities can be used for enabling CCs to efficiently com-
municate with the appropriate CHs. We note that it is also feasible to add a third-party
HTTP proxy or an open source load-balancer (e.g. Traefik1) by configuring the PEP
clients. In either way, incoming access control requests are dispatched in a balanced
way to PDP cluster nodes and the raw context is collected and handled proportionally
by the CHs. We note that any of the incoming requests to be authorized come from
authenticated entities which implies the implementation and integration of appropriate
authentication mechanisms, a work that is out of the scope of this paper.
1
https://traefik.io/.
A Context-Aware Service for Authorizing Multi-cloud Deployments 1003
are acquired from other sources (Fig. 2 – Step 2b, e.g., requestor location and device),
and (c) environment/platform-related context attributes, not pertaining to a specific
access request (Fig. 2 – Step 2a, e.g., operational status of a platform component). The
difference in the context in the two latter cases is that the request context becomes
invalid when the request has been processed, whereas environment/platform context
evolves independently of the access requests. Figure 2 provides a high-level view of
the attribute flow in the Authorization Service. Specifically, once the context is col-
lected (Fig. 2 – Steps 1, 2a, 2b), it becomes available for the PIP (Fig. 2 – Step 3),
which upon request provides them in the form of key-value pairs to the PDP for issuing
an access control decision (Fig. 2 – Step 4).
In this section, we highlight the key components of a platform dedicated for managing
applications deployments on multi-clouds that can be protected by exploiting the
implemented Authorization Service. This kind of protection is enabled by deploying
appropriate PEPs on each of the following platform components. In this way, we
consider a minimal overhead in the implementation of the developed Authorization
Service, since the approach implies that a PEP library is bundled with the application
code (i.e. same Java Archive (JAR) or placed in the Classpath), while a Tomcat
interceptor or “@AuthorizationRequired” annotations on guarded methods are added.
Business Process Management engine (BPM). It coordinates all the decision and
optimization components (i.e. Upperware components in [2]) and executes the work-
flow to generate and execute an application deployment plan, according to a set of user-
defined preferences, constraints and optimization goals. When necessary, it also repeats
the whole process or parts of it to introduce deployment plan updates, as a response to
changes in application demands or environment. A PEP client has to be embedded in
BPM engine, in order to protect it from a potentially compromised or malfunctioning
component or from outside-world interactions. In this way each PEP examines the
origin and timeliness of the requests in order to authorize them.
Deployment and Adaptation mechanism. It is responsible for taking an application
deployment plan and executing it by providing specific instructions to components able
to interface with several cloud providers that should be used, for commissioning VMs
and installing application components accordingly. In order to verify that a given
deployment plan conforms to the application deployment policies, a pre-authorization
step is taken. The plan parameters are checked against the relevant policies, and if
rendered as conformant, the deployment starts. For this reason, the Deployment and
Adaptation mechanism uses a PEP client to contact PDP to evaluate the plan against
the posed policies. The plan pre-authorization policies are different from access
authorization policies used for checking the access to previous components.
Modelling Editors. They are used to firstly create and maintain the Metadata
Schema [1], valuable for providing a formal hierarchical view of the contextual attri-
butes to be used for modelling the application placement problem. Secondly, it is used
for acquiring, from the DevOps, the appropriate application placement requirements,
constraints and optimization goals that should guide the platform when undertaking the
1004 Y. Verginadis et al.
2
http://spring.io/projects/spring-boot.
3
http://tomcat.apache.org/.
4
https://www.eclipse.org/jetty/.
A Context-Aware Service for Authorizing Multi-cloud Deployments 1005
identified via pointcuts, which are specifications of those code artefacts (typically
classes and methods) that need to be enhanced with advices. An advice, along with the
pointcuts that specify the code it must be applied onto, is called an Aspect. We note that
the Before aspects are used to enforce PEP functionality, which means that a call to a
guarded method is intercepted, PEP checks that this call is permitted and then the actual
method execution occurs.
AOP allows the non-core functionality of a software component (e.g. logging of
code executions, measuring duration, and authentication/authorization) to be moved
away from the code implementing the core business of the component. The non-core
functionality is added and interleaved with the core functionality during the software
build phase (via weaving). The Spring framework provides an AOP implementation.
Spring AOP is proxy-based, meaning each code artefact that can be enhanced with
advices will be wrapped by a suitable proxy object that is actually invoked by the
calling code. The proxy can subsequently pass control to the actually requested code.
Proxies are automatically introduced at code-level (during weaving), while source code
remains intact. Thus, this process is transparent to the programmer.
Regarding the use of Aspects in the Authorization Service, an authorization aspect
has been introduced. The corresponding advice (i.e. the wrapping proxy code) will
intercept the code invocation in order to perform a series of authorization related tasks;
namely, (a) create/reuse a PEP client object, (b) collect invocation information (i.e.
method signature and arguments), (c) connect to a PDP server and pass the collected
information, (d) receive the PDP server response (permit, deny, error), and (e) in case
of permit (subsequently), call the actual (wrapped) code or raise an authorization error,
otherwise. When the wrapped code is a method of a Web or REST controller class, and
that method is mapped to a Web or REST endpoint, then the corresponding (HTTP)
request object is introspected to extract all related information.
6 Conclusions
In terms of this work, a novel authorization service was proposed in the context of
protecting components that undertake the data-intensive application deployments on
multi-clouds. Specifically, this work is seen from two different perspectives; namely,
the “access control” to various platform components and the “pre-authorization” of
application deployment and data placement in cloud providers. In the first case,
authorization capabilities are considered to be responsible for protecting the platform
itself from illegal access attempts and interference with its normal operation. In the
latter case, authorization capabilities refer to the pre-authorization of application
deployment and data placement plans, produced by the decision-making and opti-
mization components, considering a given set of policies, constraints and goals.
The next steps of this work, involve the appropriate integration with other critical
security services that will address the security challenges in the multi-clouds deploy-
ment domain in a holistic manner. Such additional security mechanisms involve the use
of authentication services, encrypted communication (e.g. SSL/TLS), mechanisms for
counterfeiting the man-in-the-middle attacks, using component verification and digital
certificates among others.
1006 Y. Verginadis et al.
Acknowledgments. The research leading to these results has received funding from the
European Union’s Horizon 2020 research and innovation programme under grant agreement
No. 731664. The authors would like to thank the partners of the MELODIC project (http://www.
melodic.cloud/) for their valuable advices and comments.
References
1. Verginadis, Y., Patiniotakis, I., Mentzas, G.: Metadata schema for data-aware multi-cloud
computing. In: INISTA, pp. 1–9. IEEE (SMC) (2018)
2. Horn, G., Skrzypek, P.: MELODIC: utility based cross cloud deployment optimisation. In:
Proceedings of the 32nd International Conference on Advanced Information Networking and
Applications Workshops (WAINA) (2018). https://doi.org/10.1109/waina.2018.00112
3. CSA: The Treacherous 12 - Cloud Computing Top Threats in 2016. https://
cloudsecurityalliance.org/group/top-threats
4. Verginadis, Y., Michalas, A., Gouvas, P., Schiefer, G., Hübsch, G., Paraskakis, I.:
PaaSword: a holistic data privacy and security by design framework for cloud services.
J. Grid Comput. 1–16 (2017). https://doi.org/10.1007/s10723-017-9394-2
5. Decker, M.: Modelling of location-aware access control rules. In: Maria Cruz-Cunha, F.M.
(ed.), Handbook of Research on Mobility and Computing: Evolving Technologies and
Ubiquitous Impacts, Information Science Reference, Hershey, PA, USA, 2011, incollection
57, pp. 912–929 (2011)
6. Hu, V.C., Ferraiolo, D., Kuhn, R., Friedman, A.R., Lang, A.J., Cogdell, M.M., Schnitzer, A.,
Sandlin, K., Miller, R., Scarfone, K.: Guide to attribute based access control (ABAC)
definition and considerations (draft). NIST Special Publication 800, 162 (2013)
7. Ferraiolo, D., Cugini, J., Kuhn, D.R.: Role-based access control (RBAC): features and
motivations. In: Proceedings of 11th Annual Computer Security Application Conference,
pp. 241–248 (1995)
8. Ferraiolo, D., Chandramouli, R., Kuhn, R., Hu, V.: Extensible access control markup
language (XACML) and next generation access control (NGAC). In: Proceedings of the
ACM International Workshop on Attribute Based Access Control, pp. 13–24. ACM (2016)
9. Garcia-Alfaro, J., Navarro-Arribas, G.: Prevention of cross-site scripting attacks on current
web applications. In: OTM Confederated International Conferences on the Move to
Meaningful Internet Systems, pp. 1770–1784. Springer (2007)
10. Abowd, G.D., Dey, A.K., Brown, P.J., Davies, N., Smith, M., Steggles, P.: Towards a better
understanding of context and context-awareness. In: International Symposium on Handheld
and Ubiquitous Computing, pp. 304–307. Springer (1999)
11. Kiczales, G., Lamping, J., Mendhekar, A., Maeda, C., Lopes, C., Loingtier, J., M., Irwin, J.:
Aspect-oriented programming. In: European conference on object-oriented programming,
pp. 220–242. Springer (1997)
A Real-Time Video Streaming System
over IPv6+MPTCP Technology
Abstract. Today, a steadily increasing number of users are not just pas-
sively consuming Internet content, but also share and publish content.
Users publish text, photos and videos. With the availability of 5G high-
speed, low-latency mobile broadband networks, real-time video stream-
ing will also be possible. We believe this will become a very popular
application in the coming years. But the more popular a service is, the
higher the need for resilience. In this paper, we introduce our work-in-
progress live video streaming platform for future mobile edge computing
scenarios, which makes use of MPTCP+IPv6 to support multi-homing
for resilience and multi-path transport for load balancing. As a proof
of concept, we will show that the platform is (1) compatible with IPv6,
(2) utilizes load balancing when possible and (3) provides robustness by
network redundancy.
1 Introduction
With the rapid development of the Internet, more and more real-time services
are deployed in the network, such as live video broadcast, video conferencing and
other key real-time businesses. These applications have stricter requirements for
bandwidth, latency and jitter. 5G networks, together with cloud infrastructure
for mobile edge computing (MEC), are a key enabler for such applications. But in
addition, there is also need for network and transport improvements: Today, IPv4
addresses are a scarce resource, compared with new applications like the Internet
of Things (IoT). IPv6 is therefore an important development step of the next
This work has been funded by the NSFC of China (No. 61363008/61662020), CERNET
NGI Technology Innovation Project (No. NGII20160110) and the Research Council of
Norway (Forskingsrådet) (prosjektnummer 208798/F50).
The authors would like to thank Ted Zimmerman for his comments.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1007–1019, 2019.
https://doi.org/10.1007/978-3-030-15035-8_99
1008 Y. Luo et al.
generation Internet. With the rapid development of the IoT, various network
services and applications based on IPv4 will definitely migrate to IPv6 [26]. But
the single-path transmission patterns of TCP cannot handle multiple addresses
per endpoint or address changes, which are usual for IPv6. Therefore, another
milestone is Multi-Path TCP (MPTCP) [16,24], providing transport redundancy
by multi-homing without changing the existing network infrastructure. It also
supports load balancing over heterogeneous networks connections [15,17,27,30].
But MPTCP is not only useful when an endpoint is connected to multiple Inter-
net connections. [20,22] measured and compared the stability of IPv4 and IPv6
Internet paths. With IPv4/IPv6 dual-stack endpoints, MPTCP may already pro-
vide additional redundancy by just using non-congruent IPv4 and IPv6 paths.
In this paper, we introduce our work in progress on a live video streaming
service based on MPTCP and IPv6, as a proof-of-concept for a resilient service
in MEC. We furthermore present initial evaluation results.
It converts the IP address src to the string dst, with a maximum length of
size. Note, both functions only handle addr in and addr in6 – not sockaddr in
and sockaddr in6. The application therefore has to assemble the latter structures
by itself! Similar to converting strings and addresses, it is also possible to resolve
DNS names and reverse-lookup DNS names for IP addresses. This is provided
by the functions gethostbyname() (DNS name resolution, like e.g. www.nntb.no)
and gethostbyaddr() (DNS reverse lookup). Originally, these functions have only
been defined for IPv4. Most operating systems, however, provide extended IPv6-
capable functions as well. This makes platform-specific code in multi-platform
applications necessary. Furthermore, gethostbyname() and gethostbyaddr() are
not thread-safe, creating further platform-specific variants. Therefore, the usage
of these functions is now deprecated. gethostbyname() and gethostbyaddr() have
been replaced by new functions, which also handle the address/string conversion
provided by inet pton() and inet ntop(), in a platform-independent way.
The getaddrinfo() function provides the translation of an address or DNS
name – as well as a port number or service name – into IP addresses, e.g.
sockaddr in (IPv4) or sockaddr in6 (IPv6) [19,25] as necessary:
i n t g e t a d d r i n f o ( const char∗ host ,
const char∗ serv ,
const struct a d d r i n f o ∗ hints ,
struct a d d r i n f o ∗∗ res )
That is, getaddrinfo() takes an address or DNS name string as host parameter,
a port number string or service name as serv parameter, an addrinfo structure
as hints parameter, as well as a pointer res to store the result (in form of
an addrinfo structure allocated by getaddrinfo()). The corresponding structure
addrinfo is shown in Listing 1. The hints structure can provide some specific
information for the address structure to be created, particularly the ai family
(AF INET or AF INET6 – but in most cases AF UNSPEC to support any IP
protocol). Furthermore two flags in ai flags can enforce the usage of IP address
only (AI NUMERICHOST, i.e. no DNS name allowed) and port number only
(AI NUMERICSERV, i.e. no service name allowed). We will later make this
clearer with examples. The resulting addrinfo structure contains the address in
A Real-Time Video Streaming System over IPv6+MPTCP Technology 1011
ai addr (i.e. sockaddr in for IPv4 or sockaddr in6 for IPv6) and its length in
ai addrlen. If there are multiple addresses (e.g. an IPv6 and an IPv4 address),
addrinfo structures may be chained using a pointer to the next addrinfo structure
in ai next. Since getaddrinfo() allocates storage space for the resulting addrinfo
structure(s), it is necessary to free them. This is realized by calling
void f r e e a d d r i n f o ( struct a d d r i n f o ∗ r e s )
The reverse direction, i.e. address to string or DNS name, is provided by:
int getnameinfo (const struct sockaddr∗ addr, socklen t addrlen,
char∗ host, socklen t hostlen ,
char∗ serv , socklen t servlen ,
int flags )
It takes a sockaddr in (IPv4) or sockaddr in6 (IPv6) structure in addr, its length
in addrlen and flags in flags. The result is written to buffers provided by
host for address string/hostname (with maximum length hostlen) and host for
port number string/service name (with maximum length servlen). Similar to
getaddrinfo(), the flag NI NUMERICHOST turns off DNS reverse lookup (the
result will be an address string); NI NUMERICSERV turns off service name
lookup (the result will be a port number string). We will make this clearer with
examples below.
The Socket API functions to create, bind and connect sockets, as well as
to write to and receive from addresses, are using sockaddr in (for IPv4) and
sockaddr in6 (for IPv6). Since these basics are part of various publications, like
e.g. [25], we do not go into details here. To write protocol-independent code,
it is therefore useful to let getaddrinfo() create all protocol-specific information
and let getnameinfo() make this information human-readable again. Ideally, the
application itself does not have to contain any IP-protocol-specific code.
In order to make the explanations above clearer, Listing 2 shows an example
to translate an IPv6 address string and a port number string to the corresponding
sockaddr in6 structure. getaddrinfo() obtains this information, in form of an
addrinfo structure (see Listing 1), pointing to the actual sockaddr in6 structure
in ai addr.
1012 Y. Luo et al.
Listing 3. An Example for getaddrinfo() with DNS Name and Service String
Working with Internet applications, some servers are already using IPv6,
while some others still use IPv4. It is therefore convenient to use DNS names
instead of addresses. The example in Listing 3 therefore shows how to use
getaddrinfo() with DNS name (“www.nntb.no”) and service name (“https”)
instead. Calling getaddrinfo() is similar to Listing 2, except for not setting
the AI NUMERICHOST and AI NUMERICSERV flags (since we want to use
DNS and service lookup now). Since hosts may have multiple addresses in the
DNS, the example also iterates over all returned addresses chained by the
ai next pointer of the addrinfo structures. All sockaddr in and sockaddr in6
will contain the port number of service “https” (which is 443 – resolved from
/etc/services).
Finally, Listing 4 presents the usage of getnameinfo() to translate an address
(here: as result from getaddrinfo()) to a string, including the port number.
Obviously, not setting the flags NI NUMERICHOST and NI NUMERICSERV
would try to reverse-lookup the DNS name and convert the port number to a
service name.
A Real-Time Video Streaming System over IPv6+MPTCP Technology 1013
However, SRS was solely based on TCP over IPv4, i.e. without support for
IPv6. As part of the work for this paper, we added IPv6 support as well. The
changes for this support have been submitted as pull request3 to the upstream
project. It has already been merged into the upstream SRS sources, i.e. IPv6
support is now available for all SRS users. Basically, the idea of the changes
for SRS was to replace all IPv4-specific code by IP-protocol-independent code
based on getaddrinfo() and getnameinfo() (see Sect. 3). That is, getaddrinfo() is
now used to convert address strings (see Listing 2) or resolve host names (see
Listing 3) to either sockaddr in (IPv4) or sockaddr in6 (IPv6). Then, the appli-
cation code can create sockets of the right family (AF INET or AF INET6 –
returned by getaddrinfo() in ai family of the addrinfo structure; see Listing 1).
In the reverse direction, socket addresses are translated into strings by getname-
info() (see Lising 4). By making the code IP-protocol-independent, we added
IPv6 support. This not only simplified the original code, but also made it more
portable as well.
We furthermore adapted bwm-ng4 , a tool to record the data statistics of
network interfaces, with an improved accuracy of the recorded information. This
change had also been provided as pull request5 and been merged by the upstream
project.
Finally, for the video display at the client, we use ffplay. It is a playback
component of the well-known FFmpeg6 tools.
NIC
Seamless Switching
Load Balancing
NIC1
Robustness
NIC2
Time
1 2 3 4
I=NIC 1
I=NIC 2
Outging Data Rate D [KiB/s]
750
500
250
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Relative Time t [s]
I=NIC 1
Incoming Data Rate D [KiB/s]
I=NIC 2
750
500
250
0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Relative Time t [s]
References
1. Becke, M., Dreibholz, T., Adhari, H., Rathgeb, E.P.: On the fairness of transport
protocols in a multi-path environment. In: Proceedings of the IEEE International
Conference on Communications (ICC), Ottawa, Ontario, Canada, pp. 2666–2672
(2012)
2. Chen, J.L., Chao, H.C., Kuo, S.Y.: IPv6: more than protocol for next generation
internet. Comput. Commun. 29, 3011–3012 (2006)
3. Deering, S.E., Hinden, R.M.: Internet Protocol, Version 6 (IPv6) Specification.
Standards Track RFC 2460, IETF (1998)
4. Dreibholz, T.: Reliable server pooling – evaluation, optimization and extension of
a novel IETF architecture. Ph.D. thesis, University of Duisburg-Essen, Faculty
of Economics, Institute for Computer Science and Business Information Systems
(2007)
5. Dreibholz, T.: Evaluation and optimisation of multi-path transport using the
stream control transmission protocol. Habilitation treatise, University of Duisburg-
Essen, Faculty of Economics, Institute for Computer Science and Business Infor-
mation Systems (2012)
6. Dreibholz, T.: NorNet – building an inter-continental internet testbed based on
open source software. In: Proceedings of the LinuxCon Europe, Berlin, Germany
(2016)
7. Dreibholz, T.: NorNet – the internet testbed for multi-homed systems. In: Pro-
ceedings of the Multi-Service Networks Conference (MSN, Coseners). Abingdon,
Oxfordshire (2016)
8. Dreibholz, T.: A practical introduction to NEAT at Hainan University. Invited
Talk at Hainan University, College of Information Science and Technology (CIST)
(2017)
9. Dreibholz, T.: An introduction to multi-path transport at Hainan University.
Keynote Talk at Hainan University, College of Information Science and Technology
(CIST) (2017)
10. Dreibholz, T.: Big data applications on multi-clouds: an introduction to the
MELODIC project. Keynote Talk at Hainan University, College of Information
Science and Technology (CIST) (2017)
11. Dreibholz, T.: NEAT sockets API. Internet Draft draft-dreibholz-taps-neat-
socketapi-04, IETF, Individual Submission (2019)
1018 Y. Luo et al.
12. Dreibholz, T., Becke, M.: The RSPLIB project – from research to application. In:
Demo Presentation at the IEEE Global Communications Conference (GLOBE-
COM) (2010)
13. Dreibholz, T., Mazumdar, S., Zahid, F., Taherkordi, A., Gran, E.G.: Mobile edge as
part of the multi-cloud ecosystem: a performance study. In: Proceedings of the 27th
Euromicro International Conference on Parallel, Distributed and Network-Based
Processing (PDP), Pavia, Lombardia, Italy (2019)
14. Dreibholz, T., Zhou, X., Becke, M., Pulinthanath, J., Rathgeb, E.P., Du, W.: On
the security of reliable server pooling systems. Int. J. Intell. Inf. Database Syst.
(IJIIDS) 4(6), 552–578 (2010)
15. Dreibholz, T., Zhou, X., Fu, F.: Multi-path TCP in real-world setups – an eval-
uation in the NorNet core testbed. In: 5th International Workshop on Protocols
and Applications with Multi-Homing Support (PAMS), Gwangju, South Korea,
pp. 617–622 (2015)
16. Ford, A., Raiciu, C., Handley, M., Bonaventure, O.: TCP extensions for multipath
operation with multiple addresses. RFC 6824, IETF (2013)
17. Fu, F., Zhou, X., Dreibholz, T., Wang, K., Zhou, F., Gan, Q.: Performance compar-
ison of congestion control strategies for multi-path TCP in the NorNet testbed. In:
Proceedings of the 4th IEEE/CIC International Conference on Communications
in China (ICCC), Shenzhen, Guangdong, People’s Republic of China, pp. 607–612
(2015)
18. Fu, F., Zhou, X., Tan, Y., Dreibholz, T., Adhari, H., Rathgeb, E.P.: Performance
analysis of MPTCP protocol in multiple scenarios. Comput. Eng. Appl. 52(5),
89–93 (2016)
19. Gilligan, R., Thomson, S., Bound, J., McCann, J., Stevens, W.R.: Basic socket
interface extensions for IPv6. Informational RFC 3493, IETF (2003)
20. Golkar, F., Dreibholz, T., Kvalbein, A.: Measuring and comparing internet path
stability in IPv4 and IPv6. In: Proceedings of the 5th IEEE International Confer-
ence on the Network of the Future (NoF), Paris, France, pp. 1–5 (2014)
21. Gran, E.G., Dreibholz, T., Kvalbein, A.: NorNet core - a multi-homed research
testbed. Comput. Netw. Spec. Issue Future Internet Testbeds 61, 75–87 (2014)
22. Livadariu, I.A., Ferlin, S., Alay, Ö., Dreibholz, T., Dhamdhere, A., Elmokashfi,
A.M.: Leveraging the IPv4/IPv6 identity duality by using multi-path transport.
In: Proceedings of the 18th IEEE Global Internet Symposium (GI) at the 34th
IEEE Conference on Computer Communications (INFOCOM), Hong Kong, Peo-
ple’s Republic of China, pp. 312–317 (2015)
23. Parmar, H., Thornburgh, M.: Adobe’s real time messaging protocol. Technical
report, Adobe (2012)
24. Raiciu, C., Paasch, C., Barré, S., Ford, A., Honda, M., Duchêne, F., Bonaventure,
O., Handley, M.: How hard can it be? Designing and implementing a deployable
multipath TCP. In: Proceedings of the 9th USENIX Conference on Networked
Systems Design and Implementation (NSDI), San Jose, California, USA, pp. 1–14
(2012)
25. Stevens, W.R., Fenner, B., Rudoff, A.M.: Unix Network Programming. Addison-
Wesley Professional, Boston (2003)
26. Tadayoni, R., Henten, A.: From IPv4 to IPv6: lost in translation? Telemat. Inf.
33(2), 650–659 (2016)
27. Wang, K., Dreibholz, T., Zhou, X., Fu, F., Tan, Y., Cheng, X., Tan, Q.: On the path
management of multi-path TCP in internet scenarios based on the NorNet testbed.
In: Proceedings of the IEEE International Conference on Advanced Information
Networking and Applications (AINA), Taipei, Taiwan, People’s Republic of China,
pp. 1–8 (2017)
A Real-Time Video Streaming System over IPv6+MPTCP Technology 1019
28. Weinrank, F., Grinnemo, K.J., Bozakov, Z., Brunström, A., Dreibholz, T., Hurtig,
P., Khademi, N., Tüxen, M.: A NEAT way to browse the web. In: Proceedings
of the ACM, IRTF and ISOC Applied Networking Research Workshop (ANRW),
Praha, Czech Republic, pp. 33–34 (2017)
29. Wischik, D., Raiciu, C., Greenhalgh, A., Handley, M.: Design, implementation and
evaluation of congestion control for multipath TCP. In: Proceedings of the 8th
USENIX Conference on Networked Systems Design and Implementation (NSDI),
Boston, Massachusetts, USA, pp. 99–112 (2011)
30. Zhou, F., Dreibholz, T., Zhou, X., Fu, F., Tan, Y., Gan, Q.: The performance
impact of buffer sizes for multi-path TCP in internet setups. In: Proceedings of the
IEEE International Conference on Advanced Information Networking and Appli-
cations (AINA), Taipei, Taiwan, People’s Republic of China, pp. 9–16 (2017)
Towards Realistic Simulations
of Arbitrary Cross-Cloud Workloads
1 Introduction
The idea that computational power could be provided as a utility to the users
dates back to 1966 [1]. However, it was not until recently that the cloud comput-
ing systems started gaining popularity, thanks mainly to the better virtualization
support and the rise of the Internet. Over the last few years, cloud computing
has fundamentally brought a paradigm shift in computing, and the industry has
witnessed an accelerated transition from traditional cluster computing architec-
tures to highly flexible cloud-based services.
Clouds bring significant advantages over traditional cluster computing archi-
tectures including flexibility, ease of setup and deployment, high-availability, and
on-demand resource allocation [2]. In addition, a highly attractive pay-as-you-
go [3] business model of the cloud computing drives a large number of enter-
prises to migrate to cloud computing solutions. However, as it is true for any
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1020–1029, 2019.
https://doi.org/10.1007/978-3-030-15035-8_100
Towards Realistic Simulations of Arbitrary Cross-Cloud Workloads 1021
1
We use the term Cross-Cloud to refer to application deployments where multiple
cloud platforms are simultaneously used to deploy application components. The
term Multi-Cloud is also popular, though. We differentiate Multi-Cloud scenarios
from Cross-Clouds – in Multi-Clouds, applications are capable of being deployed on
different cloud platforms, but one at a time, contrary to the Cross-Cloud deployments
of application components on segregated cloud platforms at the same time.
1022 N. Mohebi and F. Zahid
1 {
2 "name": "p2.xlarge",
3 "cpus": 4,
4 "memoryInMB":62464,
5 "pricePerHour":0.900,
6 "csp": "AmazonEC2",
7 "region": "US-East(Ohio)"
8 }
3.2 Workloads
4 Evaluation
parameters defining each simulation experiment are kept constant while cloudlet
sizes are randomly chosen from a range.
We define cloud applications as a graph, G(V, E), where the set of vertices, V ,
describes self-containing application components that can be deployed on a VM
in a cloud environment. The set of edges between the applications components,
E, refers to the communication links between corresponding application compo-
nents. For the sake of simplicity, we assume that all the communication links are
bi-directional. One or more cloudlets are then executed at each application com-
ponent. Using the functionality provided in our Cross-Cloud simulation frame-
work, each of the application components can be deployed independently on a
suitable VM on any of the available cloud data centers of the CSPs. However, if
only the cheapest VMs are chosen satisfying cloudlet demands, the overhead of
the communication between geographically-dispersed data centers may outrun
the cost benefits achieved by choosing the cheapest VM options.
Mean Cost (USD)
400
200
100
20
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Component Connectivity Component Connectivity
(a) Small and Medium sized applications (b) Large sized applications
the impact of the Cross-Cloud reasoning is limited (shown as squared solid line
for Cross-Cloud Reasoner and dashed circled line for the Simple cost-based Rea-
soner). However, for the medium-sized application, as shown with crossed solid
and dash line with a plus sign for Cross-Cloud and Simple Reasoner respectively,
the impact is huge. In particular, as the connectivity between the application
components increases, the cost of the applications increases very rapidly for the
Simple Reasoner. The Cross-Cloud Reasoner on the other hand is minimally
affected by the communication links, as it considers the application connectivity
into account while provisioning the VMs for the applications. The results for
the large and very large applications are shown in Fig. 2(b). Here too, we see
that the cost of the applications when Simple cost-based reasoning is used is
higher as compared to the Cross-Cloud reasoning. The impact is greater for the
large-sized applications and with higher percentage of connectivity.
250
150
0 50
0 50
2 3 4 5 2 3 4 5
Workflow Depth Workflow Depth
250
150
0 50
0 50
2 3 4 5 2 3 4 5
Workflow Depth Workflow Depth
As described in Sect. 4.1, we used different application sizes for the workflow-
based simulations as well. The work-flow application sizes are based on the max-
imum number of application components in each of the applications. A number
of the application components are then connected in each stage. The workflow
depth defines the number of layers of workflow graphs that are defined for an
application. We present the results of the workflow-aware scheduling in Fig. 3.
The small sized applications, as shown in Fig. 3(a) are the least affected. How-
ever, still, with a workflow depth of 5, the Cross-Cloud reasoner costs about $51,
which is about 28% improvement over $71 costs of the simple reasoning. The
results for medium- and large- sized workflows are given in Fig. 3(b) and (c),
respectively. The very large applications, as shown in Fig. 3(d), benefits hugely
with the help of Cross-Cloud reasoner. As presented in the figure, the 5-layer
very large workflow on average in our simulated experiments cost around $131,
which is about 1.7 times better than the cost using the simple reasoner (which
is about $301).
5 Conclusion
In this paper, we tackled the challenge of designing a Cross-Cloud simulation
framework, CrossCloudSim-Plus, where application components are allowed to
be deployed on multiple cloud platforms. In the future, we plan to further extend
our simulation framework and include support for more complex application
design through modeling, and extend our reasoning framework to include data-
aware reasoning.
Acknowledgement. This work has received funding from the European Union’s
H2020 programme under grant agreement no. 731664 (MELODIC).
References
1. Parkhill, D.F.: The Challenge of the Computer Utility (1966)
2. Mell, P., Grance, T.: The NIST definition of cloud computing. Nat. Inst. Stand.
Technol. 53(6), 50 (2009)
3. Rappa, M.A.: The utility business model and the future of computing services.
IBM Syst. J. 43(1), 32 (2004)
4. Taherkordi, A., Zahid, F., Verginadis, Y., Horn, G.: Future cloud systems design:
challenges and research directions. IEEE Access (2018). https://doi.org/10.1109/
ACCESS.2018.2883149
5. Calheiros, R.N., Ranjan, R., Beloglazov, A., De Rose, C.A., Buyya, R.: CloudSim:
a toolkit for modeling and simulation of cloud computing environments and evalu-
ation of resource provisioning algorithms. Softw. Pract. Exp. 41(1), 23–50 (2011)
6. Silva Filho, M.C., Oliveira, R.L., Monteiro, C.C., Inácio, P.R., Freire, M.M.:
CloudSim plus: a cloud computing simulation framework pursuing software engi-
neering principles for improved modularity, extensibility and correctness. In: 2017
IFIP/IEEE Symposium on Integrated Network and Service Management (IM), pp.
400–406. IEEE (2017)
Towards Realistic Simulations of Arbitrary Cross-Cloud Workloads 1029
Simula Research Laboratory, Martin Linges vei 25, 1364 Fornebu, Norway
{dipesh,feroz}@simula.no
Over the last decade with increasing cloud popularity, centralized cloud data
center deployments have been widely used by a large number of cloud applica-
tions [1]. One major disadvantage of single data center application deployments
is that they are prone to single-point-of-failure in case of data center disrup-
tions, for instance, due to a natural disaster. On the other hand, the emergence
of the Internet-of-Things (IoT), edge computing, and over the Internet multi-
media applications demands services be provided at closest proximity to the
end-users, pushing the need of distributed deployments. To address this prob-
lem, distributed cloud data center application deployments are employed [2],
which enable cloud applications and services to be replicated at geographically-
distributed data center locations [3]. However, it is not a trivial task to select
an appropriate set of distributed cloud data centers for an application. The
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1030–1040, 2019.
https://doi.org/10.1007/978-3-030-15035-8_101
Data Center Clustering for Geographically Distributed Cloud Deployments 1031
The rest of the paper is organized as follows: Sect. 2 introduces the relevant
background, related work, and the employed clustering mechanism. Section 3
presents the experimental design, followed by the evaluation of the experiments
in Sect. 4. Section 5 describes some use cases of our data center clustering. Finally,
Sect. 6 concludes the paper.
than to the objects from a different cluster as measured by some distance func-
tion (e.g., Euclidean distance). Clustering is widely used in data mining, and the
major applications of clustering include scientific data analysis, document cate-
gorization, pattern recognition, and image processing. Clustering can be used to
obtain interesting patterns and distribution of data directly from data sets with
little or no background knowledge [5]. It is important to note that the quality of
a clustering approach is dependent on the implementation of the approach, e.g.,
the similarity measure used by the approach, selected clustering algorithm and
its required parameters (e.g., number of clusters for partitioning algorithms, such
as k-means), the definition and representation of cluster chosen by the user [6].
There are a number of existing works on data clustering across different
disciplines in the field of data mining [7]. Recently, Devi and Murugaboopathi
[3] proposed a cloud data center clustering algorithm by modeling the cloud data
center as a network graph. Specifically, in [3], distributed cloud data centers are
clustered based on their proximity by considering the degree of the cloud data
centers for cluster center selection and the distance between cloud data centers
by employing Dijkstra’s algorithm. The results were evaluated on a distributed
cloud environment simulated using the C programming language over Ubuntu
14.04 operating system.
As compared to the existing literature, we clustered data centers across dif-
ferent geographical locations from the same cloud providers and across cloud
providers. To compute the similarity between data centers, we used three dif-
ferent metrics: latency, bandwidth, and cost of data transfer. In addition, we
performed experiments on different data centers from real cloud providers for
authentic data.
To cluster the data centers, we employed affinity propagation clustering (AP)
[8] that is based on message passing techniques. Specifically, AP is an exemplar-
based clustering algorithm which finds a set of data points that best exemplify
the data and associates each data point with one exemplar. AP works based on
similarities between pairs of data points (i.e., n * n similarity matrix for n data
points). The detailed description of AP can be checked from [8]. In our context,
we used affinity clustering because AP does not require the number of clusters to
be prespecified unlike other clustering techniques (e.g., k-means [9]) and AP is
a fast clustering algorithm. In addition, we can measure the similarity between
each pair of data centers using the specified metrics (e.g., latency) as required
by AP.
3 Experimental Design
3.1 Research Questions
Our experimental design approach aims at answering the following research ques-
tions.
• RQ1. Do data centers that are close geographically perform better than data
center that are located geographically far?
Data Center Clustering for Geographically Distributed Cloud Deployments 1033
• RQ2. Can data centers from the same cloud provider in different geograph-
ical locations be clustered together based on the cost of data transfer, and
performance (latency and bandwidth)?
• RQ3. Is it more effective to use data center across different geographical
locations from the same cloud provider or across different cloud providers?
We selected 24 data centers from two popular cloud providers: Amazon Web
Services (AWS), and Google Cloud Platform (GCP). We used (1) data centers
across all the regions (i.e., 15 regions) in AWS and (2) data centers from GCP in
identical locations (i.e., 9 regions) as AWS. Specifically, we selected data centers
in N. Virginia, Ohio, N. California, Oregon, Mumbai, Seoul, Singapore, Sydney,
Tokyo, Canada (Central), Frankfurt, Ireland, London, Paris, and Sao Paulo for
AWS. In addition, we selected data centers in N. Virginia, Oregon, Mumbai,
Singapore, Sydney, Tokyo, Frankfurt, London, and Sao Paulo for GCP.
We used the tool iperf3 1 to measure the bandwidth across data centers by setting
up data center in one region as a server and the other data centers as a client.
We repeated this process for data centers in all the regions for both AWS and
GCP. In addition, we used ping to measure latency between the data centers.
Experiments were performed using iperf3 and ping for 100 s between each data
centers, and we took an average value for each pair of data center. The time 100 s
was chosen as bandwidth and latency statistics tend to converge around that time
for a given experiment. Moreover, we chose TCP to include the impact of packet
loss in our experiments. To calculate the data transfer cost for datacenter from
the same/different cloud providers, we consulted documentation from AWS2 and
GCP3 . Note that the datacenters have a different cost per data transfer usage
based on the location of data center and if it belongs to the same/different cloud
provider. In our experiments, we took the average cost of data transfer per TB.
Finally, all the values (latency, bandwidth, and cost) were normalized in the
range [0, 1].
The affinity propagation clustering algorithm was implemented from [10].
Regarding the parameter settings for AP, we used the same settings from [10]:
damping factor lambda = 0.5, convergence = 10, maximum number of itera-
tions = 1000. The implementation (using Java and Python) and datasets (i.e.,
latency, bandwidth, and cost between the data centers) are provided online4 .
All the experiments were conducted on the data centers with Ubuntu 16.04 LTS
operating system.
1
https://iperf.fr/.
2
https://aws.amazon.com/ec2/pricing/on-demand/.
3
https://cloud.google.com/compute/pricing#internet egress.
4
https://github.com/Dipsy88/M2EC Data.
1034 D. Pradhan and F. Zahid
Latency (Normalized)
Latency (Normalized)
1.0
1.0
Tokyo Mumbai
0.8
0.8
Singapore
Sydney Sydney
Seoul
0.6
0.6
Tokyo
Sao Paulo
Sao Paulo
0.4
0.4
Singapore
California
Oregon
Mumbai
Toronto
0.2
0.2
Virginia
Virginia
Dublin
0.0
0.0
London
Frankfurt London Frankfurt
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Fig. 1. Data center distances. Note Frankfurt is the base data center.
4 Evaluation
4.1 Data Center Distances
Frankfurt, Dublin, London, Mumbai, Paris Frankfurt, Dublin, London, Mumbai, Paris
Toronto, Virginia, Ohio, SaoPaulo Toronto, California, Virginia, Ohio, Oregon, SaoPaulo
250
Latency (ms)
California, Oregon, Sydney Seoul, Singapore, Sydney, Tokyo
Seoul, Singapore, Tokyo
150 Base
Base
50
50
0
0
0 50 100 150 0 50 100 150
Toronto, California, Viginia, Ohio, Oregon Toronto, California, Virginia, Ohio, Oregon
Latency (ms)
SaoPaulo, Seoul, Singapore, Sydney, Tokyo Mumbai, Singapore
Seoul, Tokyo
150
Base
Base
50
50
0
4.2 Clusters
which is better than any other inter-cluster bandwidth from the base cluster
(the second best being around 116 Mbps). As latency does not play a role in
clustering, geographically far-away data centers with expected higher latency
between them, such as Oregon and Sydney, are in the same cluster. Similarly,
the latency within one of the clusters formed is actually slightly better than
what is achieved within the base cluster. This indicates that if an application
is geographically-distributed and is not sensitive to latency and require higher
bandwidth, it is quite possible that geographically distant data centers make a
better combination than the nearby data centers. However, when the cluster-
ing employs latency as the considered metric, as shown in Fig. 2(b), clusters are
formed using data centers that are geographically close to each other. Taking
the last example, Sydney is now in a Asia-Pacific cluster, while Oregon lies in a
cluster of American data centers.
200
Virginia, Oregon, SaoPaulo Mumbai, Singapore, Sydney, Tokyo
Latency (ms)
Latency (ms)
Mumbai, Singapore, Sydney, Tokyo
150
Base
150
100
100
50
50
Base
0
Latency (ms)
Base Base
0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Figure 2(c) shows clustering based on the costs. As data transfer pricing at
AWS varies only slightly between the data centers, we see clusters formed based
on a combined weighted function of AWS charges, latency, and bandwidth with
12.5% weight for latency and bandwidth each. For instance, the data transfer
from Europe to locations in the USA is the same ($0.02 per GB), but the data
centers located there are in different clusters considering latency and bandwidth
between the data centers too. Finally, Fig. 2(d) shows clustering based on our
Data Center Clustering for Geographically Distributed Cloud Deployments 1037
fourth method where each of the cost, latency, and bandwidth are given equal
weights in the weighted function. As shown in the figure, four different clusters
are formed. Compare that with Fig. 2(a) and (b), the data center in Mumbai was
in the cluster with Frankfurt, Dublin, London, and Paris, in both cases. However,
when a weighted method is used which also considers the cost, Mumbai now
constitutes a cluster with Singapore. This type of clustering is very useful for
selecting data centers for a geographically-distributed application that follows a
utility function to assess the importance of cost and performance metrics in the
data center selection.
Our clustering results for the GCP are given in Fig. 3. Thus, we can answer
RQ2 as: it is indeed possible to cluster data centers from different geographical
locations from the same cloud provider.
250
Latency (ms)
Latency (ms)
Mumbai (AWS,Google) SaoPaulo (AWS,Google)
Virginia (AWS,Google) Mumbai (AWS,Google), Singapore (AWS,Google), Tokyo (AWS,Google)
Oregon (AWS,Google) Sydney (AWS,Google)
SaoPaulo (AWS,Google)
Singapore (AWS,Google)
Sydney (AWS,Google)
150
150
50
50
Base Base
0
Latency (ms)
50
50
Base Base
0
Fig. 4. Clusters when data centers from both AWS and GCP are used
centers from Frankfurt and London) is about 400 Mbps, which is more than
four times better than the average bandwidth achieved from the base cluster to
the nodes in the data center of the second best cluster (Virginia). Latency also
follows a similar trend. Moreover, the clusters based on both the bandwidth and
latency, as shown in Fig. 4(a) and (b), respectively, indicate that geographical
location plays a major role in Multi-Cloud data center cluster formation. That
is, the data centers which are in the same city ends up in the same cluster,
irrespective of the cloud provider. Actually, the same trend can be observed on
the clusters formed using all four of our clustering methods, including the cost-
based clusters shown in Fig. 4(c). This is because of the following reasons. First,
when talking about the available bandwidth, even though one can expect that
the bandwidth available between the data centers of the same cloud provider
would be better than the bandwidth available across cloud providers, this is
not particularly true. Thus, we can answer RQ3 as: Available bandwidth across
geographically-dispersed locations is only marginally affected by the choice of the
cloud provider. Similarly, the latency also does not vary considerably based on
the cloud provider, given it is largely determined by the geographical location of
the data centers. For the AWS, the cost of data transfer from AWS data centers
to outside costs considerably higher than within AWS data centers. However, for
the GCP, the charges are the same irrespective of whether the data is transferred
to their own data centers or outside GCP. With the multi-cloud setup using
GCP and AWS, GCP uniform pricing dampens the impact of AWS pricing on
the clusters obtained.
top of it, cost is always a concern in the cloud computing. Employing the data
center clustering mechanism we presented in this paper, available cloud data
centers can be clustered based on the application requirements, yielding most
cost-efficient distributed application deployment. Furthermore, our experimen-
tation also gave an insight into the available performance and cost characteristics
between data center of two major cloud providers.
6 Conclusion
Acknowledgement. This work has received funding from the European Union’s
H2020 programme under grant agreement no. 731664 (MELODIC).
References
1. Liu, Y., Muppala, J.K., Veeraraghavan, M., Lin, D., Hamdi, M.: Data Center
Networks: Topologies, Architectures and Fault-Tolerance Characteristics. Springer,
Heidelberg (2013)
2. Khethavath, P., Thomas, J.P., Chan-tin, E.: Towards an efficient distributed cloud
computing architecture. Peer-to-Peer Netw. Appl. 10(5), 1152–1168 (2017)
1040 D. Pradhan and F. Zahid
1 Introduction
Humans have always been concerned with managing our environments, and
research has made tremendous progress in complexity and cybernetics over the
last century [26]. Approaches based on Model Predictive Control (MPC) have
become the standard way of successfully solving constrained optimal control prob-
lems. However, nature is stochastic, and solving optimal control problems under
uncertainty suffers from the ‘curse of dimensionality’ as the inherent Dynamic
Programming (DP) problem grows exponentially with the dimensionality. In this
case, Stochastic Model Predictive Control (SMPC) [3] has recently appeared as
an alternative, or one must resort to approximation strategies [15] based on Rein-
forcement Learning (RL) [21] or Neuro-Dynamic Programming (NDP) [8], which
uses Neural Networks (NNs) [25] to approximate the objective function.
This work has received funding from the European Union’s Horizon 2020 research
and innovation programme under grant agreement No 731664 MELODIC: Multi-cloud
Execution-ware for Large-scale Optimised Data-Intensive Computing, and from the
European Union’s EUREKA Eurostars research and innovation programme under
agreement No E! 11990 FUNCTIONIZER: Seamless support of serverless applications
in multi-cloud.
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1041–1054, 2019.
https://doi.org/10.1007/978-3-030-15035-8_102
1042 G. Horn et al.
sets of input data. Furthermore, one may need many attempts using different
approaches to find the best method for pre-processing data. Pre-processing of
data and training require significant computing power, but usually the execution
time is not strictly limited.
However, the main objective of many CI systems is to generate predictions
in real-time or near real-time [1]. Good examples are the applications discussed
in Sect. 3: financial applications used for investment portfolio optimization, and
dynamic pricing model for vehicle-sharing operators; and many others like inter-
net advertisement or e-commerce retail recommender systems. The predictions
must be calculated very fast and reliably as delayed calculations make the results
obsolete or useless.
Thus, massive computing power is needed for CI applications, and the calcu-
lations are usually done in a distributed way, owing to the high cost of High Per-
formance Computing (HPC) solutions. Cloud resources are usually less expensive
than HPC and more flexible to configure and set up than a HPC infrastructure.
The Cloud computing model allows for very flexible creation and sizing of the
application’s computing infrastructure based on available budget and the Cloud
providers’ offers.
On the other hand, a distributed architecture is usually more complex and
difficult to manage and to optimize than monolithic architectures, and proper
use of cross-Cloud computing, and maximizing its efficiency, are the most impor-
tant challenges for distributed CI applications. A proper computing architec-
ture design with enough performance and reliability is crucial for these applica-
tions. Owing to the distributed architecture, many Cloud providers offer different
options and pricing, and so choosing the optimal solution is very difficult. The
cost savings and performance gains, in terms of number of trained models, pre-
processed data, and number of simulations to run, could be really significant.
However, continuously deploying, adapting, and decommissioning the pro-
vided Cloud resources manually will be challenging. The Multi-cloud Execu-
tion ware for Large scale Optimised Data Intensive Computing (MELODIC)
platform1 offers utility based model driven autonomic cross-Cloud deployment
while respecting the application’s constraints on data placement, timeliness, and
cost [11].
Extracting value from available data is becoming the key differentiator in the
financial market. Systems using Artificial Intelligence (AI) can analyse much
more data than humans, and therefore they are expected to react quicker and
make better investment decisions based on more information. The start-up AI
Investments 2 uses AI in a self-improving platform for optimizing an investment
1
http://www.melodic.cloud/.
2
http://www.aiinvestments.pl/.
1044 G. Horn et al.
portfolio in a complete trading solution for the global markets, including a diver-
sified way of signalling transactions, determining market conditions, and man-
aging exposition. Instead of predicting the prices of portfolio assets, the solution
offers “diversified high probability transaction notifications” for exposure, risk,
and position sizing management by applying deep CNN, Long Short-Term Mem-
ory networks [17], as well as an reinforcement learning algorithm based on Monte
Carlo Tree Search with value network [24].
Hence, the application is fully distributed, and both the preprocessing data step
and the training step could be done in a distributed manner. The computing
requirements for this application are:
1. The need for training many Machine Learning (ML) models for pattern recog-
nition, usually with complex architectures [20]. Currently, AI Investments
trains a dedicated model for each market, for each time interval (hourly, daily,
weekly), and for each investment strategy. It means that for 200 markets, 3
time intervals and 5 investment strategies, 3000 models must be trained and
re-trained periodically based on new data.
3
https://kafka.apache.org/.
Cost Benefits of Multi-cloud Deployment 1045
4. The MCTS component performs the MCTS roll-out to find the best set of
relocations. For each roll-out, NNs are used to predict the profitability and
the degree of satisfied customer demand.
5. The GA optimisation component performs the GA based optimization to
find the best set of relocations. For each generation, NNs are used to predict
the profitability and the degree of satisfied customer demand. Alternatively
simulations are performed.
6. Vehicle Routing Problem Solver is used to find optimal routes for the ser-
vice work force to perform neccesary service activities such as refueling or
relocation.
7. Additional worker components for NNs training. These are used for train-
ing each NN model for predicting profitability, and a second NN model for
predicting satisfied demand.
Simulation is the key element of the platform, because the results of the simu-
lation are used for NN training that are further used by the optimization algo-
rithms to find the optimal relocations of the vehicles.
The approach to deployment of the Optimali application is based on similar
deployment utility as for AI Investments application. The goal is to minimize
cost of resources, aiming to finish the process of simulation and prediction in the
given time frame. In this case, the utility based Cloud deployment has significant
impact on the overall work flow. Owing to MELODIC’s dynamic resource alloca-
tion by adding or removing resources, it is possible to meet the time constraints
with the optimal cost almost regardless of the size of the problem instance.
4 Discussion
Reduced cost is the main utility of both applications, and the evaluation is
therefore based on the economic benefits of a cross-Cloud deployment combining
public servers with the alternative scenario of only using private infrastructure
to fulfil the applications’ training requirements.
For the AI Investments application it is assumed that on average 10 models
are trained in parallel, but at peak level up to 50 models could be trained, and
the average number of models can reach 18 models during 120 h per month. The
distribution of the training time is presented on Fig. 3. It presents the percentage
of the models trained by a given time. The figure shows that training time is not
constant and depends of the size of the model, training period length, and the
performance of the server and GPU.
With these parameters two different scenarios were considered.
1. A scenario with 5 on-premise servers each with two GPUs. This corresponds
to two models to be trained per server in the normal average case. Additional
servers with GPUs running in the Cloud must therefore be used as needed.
Based on above assumptions the three years’ Total Cost of Ownership (TCO)
is composed of the following elements:
1048 G. Horn et al.
Fig. 3. Training time distribution for the financial trading solution where a few com-
plicated models require significant time to train and the average training time is 7.1 h.
a. The initial purchase of 5 on-premise servers where each server costs 5 000
USD, giving a total investment of 25 000 USD in total.
b. The cost of running a server and its maintenance covering electricity,
cooling, and repairs is assumed to be 10% of the hardware cost yearly.
Thus the total cost of infrastructure operation is 7 500 USD over the
period of three years’ server depreciation.
c. Cloud resources are used only when needed. Up to 10 models trained
at once could be handled by the private infrastructure, and only if it is
necessary to train more models will the Cloud resources be used. Based on
the parameters described above, the Cloud resources are needed during
the 120 h per month when the average load is higher than normal. On
average 8 additional servers are needed during this time since the Cloud
servers only have one GPU each. The cost of such a Cloud server is around
one USD per hour, so the cost of the used Cloud resources will be 960
USD per month, and 34 560 USD over three years5,6,7 .
Adding the cost elements listed above gives a TCO of 67 060 USD for the
three years of operation. The real cost will be lower as the additional servers
in the Cloud are started on demand by the MELODIC platform and stopped
when no longer needed, hence, the extra servers will be running only when
number of trained models exceed the normal average supported by the private
infrastructure.
2. A scenario with 25 private servers with two GPUs. This corresponds to two
models to be trained per server at peak need. The cost of these servers is
162 500 USD. This is composed of 125 000 USD in hardware cost for the 25
servers, and maintenance cost assumed to be 10% percent of the hardware cost.
5
https://aws.amazon.com/ec2/pricing/on-demand/.
6
https://azure.microsoft.com/pl-pl/pricing/.
7
https://cloud.google.com/compute/pricing.
Cost Benefits of Multi-cloud Deployment 1049
The difference between the two scenarios is 95 400 USD, so the first scenario
could save almost 60% of the cost of running the application on a private
infrastructure. There are many possible Virtual Machine (VM) types or flavours
provided by the various Cloud Providers. Properly choosing the most optimal
flavour for specific task requires additional calculation. It is done automatically
by MELODIC. The list of the most suitable VM flavours for AI Investments is
depicted in Table 1.
Table 1. Suitable Cloud Virtual machine types from different Cloud providers for the
NN training done by the Financial Trading Solution with prices as available at the
time of writing. Memory is in gigabytes.
Table 2. Different ways to support the parallel training of 10 models in the normal
situation, with occasional bursts of parallel training of 50 models to be trained four
times per month. The cost is in USD.
batch i. It will have to wait for the processing of all earlier batches, which is
on average (i − 1)t̄ where t̄ is the average training of one model. This delay is
suffered by all B models in the same batch, hence the total waiting time for all
models in this batch is (i − 1) · t̄ · B. The total waiting time suffered by all models
under the batched scenario results from adding the waiting times for each of the
n batches together:
n n
1
W = (i − 1) · t̄ · B = t̄ · B (i − 1) = t̄ · B · n(n − 1) (1)
i=1 i=1
2
which follows from the sum of the arithmetic series. Table 3 contains the results
of the cost analysis of the batched processing. It contains only the number of
on premise servers in range [5, . . . , 10], because for the lower number of the on
premises servers would not be used to train models in peak time, they would be
used to train base models.
Similar to the AI Investments, also the Optimali application considers low
cost as the main utility, and the cost of using only a private infrastructure is
compared with running the same number of simulations using virtual machines
in the Cloud at on-demand prices. For the Optimali application it is assumed
that a total of 128 cores are needed to complete the simulation, training, and
optimization work by the deadline. These cores are needed up to two hours
per day, because the optimization of the car sharing fleet is done once per day
and should be done in two hours’ long time slots. With these parameters the
calculations for the two scenarios are as follows.
Cost Benefits of Multi-cloud Deployment 1051
Table 3. Splitting the models to be trained into batches and executing the jobs in one
batch in parallel, but the batches sequentially may lead to less Cloud servers used, but
also extended time to completion.
1. The scenario with only dynamic Cloud servers: The average cost of these
servers is around 5.5 USD per hour. This is based on using 8 virtual machines
with 16 cores each, where each machine cost 0.68 USD per hour. The monthly
cost of these cloud servers is therefore 330 USD based on 5.5 USD per hour,
two hours per day, and 30 days per month. The TCO over three years is
therefore 11 880 USD, which is 330 USD times 36 months.
2. The scenario using a private infrastructure requires optimally an initial pur-
chase of 11 servers with 12 cores each, as this is the most cost-efficient server
type. The cost of each server is 3 000 USD, and the total cost of the hardware
purchase is 33 000 USD. The additional cost of maintenance is again assumed
to be 10% of the hardware cost, thus amounting to 9 900 USD during the
three years. The TCO is therefore 42 290 USD for the private infrastructure.
The difference between the two scenarios is 31 020 USD, so the Cloud scenario
optimized by MELODIC could save over 70% of the cost of running the appli-
cation on a private infrastructure.
The additional benefits from using automatic cloud deployments with time
constrained utility come from the fact that optimization problems of various size
could be solved in virtually the same time due to the dynamic Cloud scalability
offered by MELODIC and its ability to start as many virtual machines as needed.
The Optimali applications may use many possible VM types or flavours, and
so properly choosing the most optimal flavour requires additional calculations,
similar to the ones shown above for the AI Investments application.
As a next steps of our research, we will investigate other Cloud computing
models. One of the most promising approach is to focus on serverless approach.
1052 G. Horn et al.
5 Conclusion
References
1. Agrawal, A., Gans, J., Goldfarb, A.: Prediction Machines: The Simple Economics
of Artificial Intelligence. Harvard Business Review Press, Brighton (2018)
2. Zhou, A., Qu, B.Y., Li, H., Zhao, S.Z., Suganthan, P.N., Zhang, Q.: Multiobjective
evolutionary algorithms: a survey of the state of the art. Swarm Evol. Comput.
1(1), 32–49 (2011). https://doi.org/10.1016/j.swevo.2011.03.001
3. Mesbah, A.: Stochastic model predictive control: an overview and perspectives for
future research. IEEE Control Syst. 36(6), 30–44 (2016). https://doi.org/10.1109/
MCS.2016.2602087
4. Bede, B.: Mathematics of Fuzzy Sets and Fuzzy Logic. Studies in Fuzziness and
Soft Computing, vol. 295. Springer, Berlin (2013)
5. Chui, C.K.: Wavelets: A Tutorial in Theory and Applications. First and Second
Volume of Wavelet Analysis and Its Applications (1992)
6. Cochran, W.T., Cooley, J.W., Favin, D.L., Helms, H.D., Kaenel, R.A., Lang,
W.W., Maling, G.C., Nelson, D.E., Rader, C.M., Welch, P.D.: What is the fast
fourier transform? Proc. IEEE 55(10), 1664–1674 (1967). https://doi.org/10.1109/
PROC.1967.5957
7. Fogel, D.B.: What is evolutionary computation? IEEE Spectr. 37(2), 26–32 (2000).
https://doi.org/10.1109/6.819926
8. Bertsekas, D.P., Tsitsiklis, J.N.: Neuro-Dynamic Programming. Optimization and
Neural Computation Series. Athena Scientific, Belmont (1996)
8
https://www.melodic.cloud/functionizer.html/.
Cost Benefits of Multi-cloud Deployment 1053
9. Dreibholz, T., Mazumdar, S., Zahid, F., Taherkordi, A., Gran, E.G.: Mobile edge as
part of the multi-cloud ecosystem: a performance study. In: Proceedings of the 27th
Euromicro International Conference on Parallel, Distributed and Network-Based
Processing (PDP), Pavia, Lombardia/Italy (2019)
10. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
CoRR abs/1512.03385 (2015)
11. Horn, G., Skrzypek, P.: MELODIC: utility based cross cloud deployment optimisa-
tion. In: 32nd International Conference on Advanced Information Networking and
Applications (AINA) Workshops, Poland, pp. 360–367. IEEE Computer Society,
Krakow (2018). https://doi.org/10.1109/WAINA.2018.00112
12. Hsieh, T.J., Hsiao, H.F., Yeh, W.C.: Forecasting stock markets using wavelet trans-
forms and recurrent neural networks: an integrated system based on artificial bee
colony algorithm. Appl. Soft Comput. 11(2), 2510–2525 (2011). https://doi.org/
10.1016/j.asoc.2010.09.007. The Impact of Soft Computing for the Progress of Arti-
ficial Intelligence
13. Witten, I.H., et al.: Data Mining: Practical Machine Learning Tools and Techniques
(2016)
14. Kacprzyk, J., Pedrycz, W. (eds.): Springer Handbook of Computational Intelli-
gence. Springer Handbooks. Springer, Berlin (2015)
15. Lee, J.M., Lee, J.H.: Approximate dynamic programming strategies and their appli-
cability for process control: a review and future directions. Int. J. Control Autom.
Syst. 2(3), 263–278 (2004)
16. Zhang, J., Zhan, Z.H., Lin, Y., Chen, N., Gong, Y.J., Zhong, J.H., Chung, H.S.H.,
Li, Y., Shi, Y.H.: Evolutionary computation meets machine learning: a survey.
IEEE Comput. Intell. Mag. 6(4), 68–75 (2011). https://doi.org/10.1109/MCI.2011.
942584
17. Kim, J., El-Khamy, M., Lee, J.: Residual LSTM: design of a deep recurrent archi-
tecture for distant speech recognition. CoRR abs/1701.03360 (2017)
18. LeCun, Y., Bengio, Y.: Convolutional networks for images, speech, and time series.
In: Arbib, M.A. (ed.) The Handbook of Brain Theory and Neural Networks, pp.
255–258. MIT Press, Cambridge (1995)
19. Chen, X.-W., Lin, X.: Big data deep learning: challenges and perspectives. IEEE
Access 2, 514–525 (2014)
20. Abadi, M., et al.: TensorFlow: a system for large-scale machine learning. In: 12th
USENIX Symposium on Operating Systems Design and Implementation (OSDI
2016) (2016)
21. Sutton, R.S., Barto, A.G.: Reinforcement Learning, vol. 9. MIT Press, Boston
(1998)
22. Eberhart, R., Shi, Y., Kennedy, J.: Swarm Intelligence. The Morgan Kaufmann
Series in Artificial Intelligence, 1st edn. Morgan Kaufmann, Burlington (2001)
23. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S.,
Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T.P., Leach, M.,
Kavukcuoglu, K., Graepel, T., Hassabis, D.: Mastering the game of Go with deep
neural networks and tree search. Nature 529, 484–489 (2016)
24. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A.,
Hubert, T., Baker, L.R., Lai, M., Bolton, A., Chen, Y., Lillicrap, T.P., Hui, F.,
Sifre, L., van den Driessche, G., Graepel, T., Hassabis, D.: Mastering the game of
go without human knowledge. Nature 550, 354–359 (2017)
1054 G. Horn et al.
25. Walczak, S.: Artificial neural networks. In: Mehdi Khosrow-Pour, D.B.A. (ed.)
Advanced Methodologies and Technologies in Artificial Intelligence, Computer
Simulation, and Human-Computer Interaction, vol. 1, pp. 40–53 (2019). Chapter 4
26. Umpleby, S.A., Dent, E.B.: The origins and purposes of several traditions in sys-
tems theory and cybernetics. Cybern. Syst. 30(2), 79–103 (1999). https://doi.org/
10.1080/019697299125299
27. Sun, S.: A survey of multi-view machine learning. Neural Comput. Appl. 23(7),
2031–2038 (2013). https://doi.org/10.1007/s00521-013-1362-6
28. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.E., Anguelov, D., Erhan,
D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. CoRR
abs/1409.4842 (2014)
29. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser,
L., Polosukhin, I.: Attention is all you need. CoRR abs/1706.03762 (2017)
30. Pedrycz, W., Ekel, P., Parreiras, R.: Fuzzy Multicriteria Decision-Making: Models,
Methods and Applications. Wiley, Hoboken (2010)
31. Pawlak, Z.: Rough sets. Int. J. Comput. Inf. Sci. 11(5), 341–356 (1982). https://
doi.org/10.1007/BF01001956
An Overview of Multi-cloud Computing
contains the physical hardware including servers, switches, power, cooling sys-
tems, etc. Usually, the Hardware Layer takes the form of data centers harbour-
ing thousands of servers. The next level up is the Infrastructure Layer, which
operates by using virtualisation technologies to partition compute and storage
resources. The Platforms Layer is made up of operating systems and application
frameworks, and is mostly used to decrease the strain of deploying applications
directly into virtual machine containers. Finally, the Application Layer consists
of the cloud computing applications, which refers to the applications (SaaS) that
are built on the lower levels (IaaS and PaaS) of the cloud network. For example,
the applications that are available on Google Play or other app stores deliver
applications that run some of their features in clouds, allowing both ease of access
and lowering the computational cost of the app on the device, as operations can
be performed in the cloud.
It is important to discuss deployment models, and overall there are four ways
in which cloud computing can be deployed. In the private model, the cloud is
provisioned for use by a single organisation with multiple users (e.g. business
consumers). The organisation may own and oversee the system, rely on a third
party for service or some combination of the two. In the community cloud model,
multiple organisations may be responsible for cloud provisioning. As with private
cloud computing, one or more of the organisations using the community cloud
may own and operate it, rely on third parties, or again, some combination of
them. A public cloud is one that is accessible by the general public and, unlike the
previous two types of cloud computing, they are only available on the premises
of the cloud provider (e.g. university campus), while the others may be accessible
off premises. Finally, hybrid cloud systems incorporate two or more of the above
deployment models which remain separate entities. However, they are unified by
an underlying technology that allows for data application portability between
the units [25].
Now that the basics of cloud computing have been reviewed, it is possible to
discuss some of the issues facing this technology. Though cloud computing is
useful to businesses as it removes the need for planned provisioning and grants
small businesses the ability to scale their computing needs growing in line with
their business requirements, there are some issues and challenges facing cloud
systems that could be the subject of future research [26]. Though still in its
infancy, the widespread use of cloud computing has allowed many researchers
and businesses to begin to weigh in on potential issues facing the technology.
Security: One subject of concern cited by many authors is that of data security
and privacy [6,15,26,27]. In a study by [15], security concerns were flagged as a
key issue in 66 research papers reviewed, with the next highest issue being infras-
tructure at 46, followed by data management at 15. In this study, practitioners
were also interviewed, expressing additional concerns, however still indicated
1058 J. Hong et al.
that security was a practical issue, highlighting that this is a pressing matter for
many researchers and businesses.
Security issues can be subdivided into several subcategories, including: safety
mechanisms, cloud server monitoring, data confidentiality, and avoiding mali-
cious operations [27]. The authors describe problems pertaining to data stor-
age in cloud computing networks, separating it into the components of data
integrity, data confidentiality, data availability, and data privacy. Maintaining
data integrity involves ensuring that data present within a system cannot be
deleted, modified, or fabricated without authorisation. Data confidentiality is
another issue as currently it can be difficult to protect against both insider
threats and external breaches limiting the utility of cloud networks for sensitive
data storage, such as medical records or government files [33]. Another dimen-
sion of security is data availability, or the degree to which data can be recovered.
Finally, data privacy refers to the ability of an individual or group to privately
and selectively share information only amongst themselves.
Despite these concerns, [6] argues that – given the proper preparation – cloud
computing may actually become more secure than other methods. The authors
explain that any obstacles facing cloud computing are also present in traditional
systems and can be solved using technologies that are already used, such as data
encryption, virtual local area networking, firewalls, and more.
Legal Concerns: Legal issues, while related to security issues, are another
distinct concern [15,27]. Data storage rules and laws vary depending on the
location [6]. Although service level agreements between providers and consumers
have been established, there are currently no standards in place. Adopting cloud
computing storage rules to fit local law may be a challenge in the future, and
will require careful planning and coordination when designing these systems.
Data Management: Perhaps an obvious concern to many is what happens to
the data on cloud computing systems in the event of a catastrophic loss of data.
If, for example, a cloud provider company were to go bankrupt, the data may
become irretrievable. This would present a huge problem for large organisations
that are dependent on data storage [15,33].
Interoperability: [23] considers interoperability, or the “ability of diverse sys-
tems and organisations to work together (inter-operate)” to be the second-
greatest challenge to cloud computing systems after data security and trust
issues. This is crucial, because it would allow users to avoid vendor lock-in,
whereby the user becomes dependent on the CSP because they cannot extract
their services to other platforms or clouds. The lack of standardisation means
that it can be difficult to transfer data between cloud systems.
High Latency: Another issue facing cloud computing is the problem of high
latency, or a lag from when the transfer of data begins following an instruc-
tion for its transfer. In the context of cloud computing, this is caused by the
necessity of the various nodes in the cloud to communicate with each other.
Two solutions have emerged as potential solutions to this problem: fog comput-
ing and edge computing. Fog computing may be defined as “a scenario where
An Overview of Multi-cloud Computing 1059
3 Multi-cloud Solutions
To help deal with some of the issues mentioned above, the next logical progres-
sion in cloud computing is in multi-cloud computing, or cloud systems that utilise
numerous cloud networks and services simultaneously [13]. In short, multi-cloud
systems simply use more than one CSP, though they come in many subcate-
gories that will be described below. This is similar to the definition of cross-cloud
architectures, defined as systems that span across multiple provisioning bound-
aries [14]. Though the exact distinction between multi-clouds and cross-clouds
is lacking (as described in further detail below), it can be understood that in
multi-clouds, the user or business in question utilizes different cloud services for
different applications in their business. For example, they may store data on a
private cloud, share documents on the Google Cloud platform, and perform data
analysis on yet another cloud. On the other hand, cross-cloud architectures are
designed to make the transfer of data and utilisation of apps across the clouds
more streamlined and cohesive [14].
There are various rationales for using multi-clouds or cross-clouds, many of
which address some of the issues mentioned in the previous section. According
to [24], multi-cloud computing directly addresses 10 key issues, some of which
have been mentioned above. Furthermore, it is important to realise that these
issues are both shared and specific to the parties involved, including individual
customers, company level customers, and the CSPs. The 10 key issues addressed
by multi-cloud computing are:
for privacy [22]. Other approaches involve divvying up data amongst clouds
for rapid data storage. In Hadoop, an open source implementation of Google’s
method for data storage, large volumes of data are cut into more manageable
chunks across thousands of computers. A parallelised programming API is then
used to distribute the computations to where the data is located and to aggre-
gate the results. This method has tremendous application for bioinformatics and
genomics. However, the technical savvy and lack of bioinformatics tools that run
in parallel has proven to be a barrier up until very recently [4]. They reviewed var-
ious big data cloud computing frameworks including Hadoop, Spark and Flink.
It will be interesting to see what further improvements will be made and how
multi-cloud computing architectures will integrate these approaches. Finally [29]
also mentions the use of Amazon Elastic Map Reduce, IBM BigInsights, and
Microsoft Azure HDInsight as methods for allowing accessible large-scale data
processing frameworks, namely Apache Hadoop, and Apache Spark.
Machine Learning: Although related to big data, machine learning presents its
own sets of challenges and applications for use with cloud computing technolo-
gies. Given the availability of data stored in clouds (e.g. photographs and videos),
there is a wealth of information that can be processed by machine learning algo-
rithms. Techniques such as virtual machine optimisation for healthcare services
are being trialed and studied, and show promise, but must be trialed with other
diseases to demonstrate consistency [1]. [16] has explored the use of hybrid cloud
storage and machine learning for use in the oil and gas industry, where techni-
cal documents contain valuable information from disciplines like geoscience and
engineering and are in general stored in a unstructured format. To improve data
extraction and utilisation, the authors propose a machine-learning-enabled plat-
form, consisting of a carefully selected sequence of algorithms, developed as a
hybrid cloud container that automatically reads and understands the technical
documents with little human supervision. The user can upload raw data to the
platform, which is stored on a private local server. Structured data is generated
as output, which is pushed through to a search engine that is accessible to the
user via the cloud. This allows for a user to quickly identify the most important
parts of technical documents, automate the extraction or relevant data from the
documents, present the data in a meaningful way for further analysis, and finally,
allows the user to share it easily and port it to other platforms.
Future directions for machine learning and multi-cloud computing research
include a shift towards the wide-scale adoption of auto-tuners, especially for the
SaaS layer of the Cloud. [10] also anticipate the advent of new automated tools
for cloud users to benefit from the experience of other users through partially
automated application builders, automated database sharers, query optimisers,
or smart load balancers and service replicators. In other words, the interface
between cloud, application, and user, is expected to become more streamlined
and intuitive based on what is learned from the user experience. As security
has always been a key concern for cloud networks of any kind, novel machine
learning methods may be used for increased security measures. However, the
authors did not specify what these may look like or on what they would be
based, representing a future area of research application.
1066 J. Hong et al.
5 Conclusion
Cloud computing is a blossoming technology with numerous applications in many
industries, including remote computing and storage. Though vendor lock-in and
cyber security are major concerns hybrid clouds, multi-clouds, and federation
clouds may address some of these problems by providing users with alternatives
in the case of scheduled maintenance, breaches, or shut-downs, though each has
their own benefits and disadvantages. Hybrid systems are easily customised to
a given application, however less transferable, and often used for only one task
whereas multi-clouds and federated clouds are suited for businesses that require
multiple tasks or services. Future work should incorporate multi-cloud paradigms
and combine them with other technologies such as machine learning and big data,
as these technologies can either be used to solve some of the current issues with
cloud computing, such as increasing security, or can be used for entirely new
methods of analysis.
References
1. Abdelaziz, A., Elhoseny, M., Salama, A.S., Riad, A.M.: A machine learning model
for improving healthcare services on cloud computing environment. Measurement
119, 117–128 (2018)
2. Abdo, J.B., Demerjian, J., Chaouchi, H., Barbar, K., Pujolle, G.: Broker-based
cross-cloud federation manager. In: 8th International Conference for Internet Tech-
nology and Secured Transactions (ICITST 2013), London, United Kingdom (2013)
3. Aggarwal, R.: Resource provisioning and resource allocation in cloud computing
environment. Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol. 3(3), 1040–1049
(2018)
4. Alzain, M.A., Soh, B., Pardede, E.: Big data, Hadoop and cloud computing in
genomics. J. Biomed. Inform. 46(5), 774–781 (2013)
5. AlZain, M.A., Soh, B., Pardede, E.: A survey on data security issues in cloud
computing: from single to multi-clouds. J. Softw. 8(5), 1068–1078 (2013)
6. Armbrust, M., Fox, A., Griffith, R., Joseph, A.D., Katz, R.H., Konwinski, A.,
Lee, G., Patterson, D.A., Rabkin, A., Stoica, I., Zaharia, M.: Above the clouds: a
berkeley view of cloud computing. Technical reports UCB/EECS-2009-28, EECS
Department, University of California, Berkeley (2009)
7. Awasthi, P., Mittal, S., Mukherjee, S., Limbasiya, T.: A protected cloud compu-
tation algorithm using homomorphic encryption for preserving data integrity. In:
Recent Findings in Intelligent Computing Techniques, Singapore (2019)
8. Benzekki, K., Fergougui, A.E., Alaoui, A.E.B.E.: A secure cloud computing archi-
tecture using homomorphic encryption. Int. J. Adv. Comput. Sci. Appl. 7(2), 293–
298 (2016)
9. Buyya, R., Son, J.: Software-defined multi-cloud computing: a vision, architectural
elements, and future directions. In: Proceedings of the 18th International Confer-
ence on Computational Science and Applications (ICCSA) (2018)
10. Buyya, R., Srirama, S.N., Casale, G., Calheiros, R., Simmhan, Y., Varghese, B.,
Gelenbe, E., Javadi, B., Vaquero, L.M., Netto, M.A.S., Toosi, A.N., Rodriguez,
M.A., Llorente, I.M., Vimercati, S.D.C.D., Samarati, P., Milojicic, D., Varela,
C., Bahsoon, R., Assuncao, M.D.D., Rana, O., Zhou, W., Jin, H., Gentzsch,
W., Zomaya, A.Y., Shen, H.: A manifesto for future generation cloud computing:
research directions for the next decade. ACM Comput. Surv. 51(5), 105 (2018)
An Overview of Multi-cloud Computing 1067
11. Dreibholz, T.: Big data applications on multi-clouds: an introduction to the melodic
project. Keynote Talk at Hainan University, College of Information Science and
Technology (CIST) (2017)
12. Dreibholz, T., Mazumdar, S., Zahid, F., Taherkordi, A., Gran, E.G.: Mobile edge as
part of the multi-cloud ecosystem: a performance study. In: Proceedings of the 27th
Euromicro International Conference on Parallel, Distributed and Network-Based
Processing (PDP), Pavia, Lombardia, Italy (2019)
13. Elkhatib, Y.: Mapping cross-cloud systems: challenges and opportunities. In: Pro-
ceedings of the 8th USENIX Conference on Hot Topics in Cloud Computing, Berke-
ley, United State (2016)
14. Elkhatib, Y., Blair, G.S., Surajbali, B.: Experiences of using a hybrid cloud to
construct an environmental virtual observatory. In: Proceedings of the 3rd Inter-
national Workshop on Cloud Data and Platforms, Prague, Czech Republic (2013)
15. Ghanam, Y., Ferreira, J., Maurer, F.: Emerging issues and challenges in cloud
computing - a hybrid approach. J. Softw. Eng. Appl. 5(11A), 923 (2012)
16. Hernandez, N.M., Lucañas, P.J., Graciosa, J.C., Mamador, C., Caezar, L., Pan-
ganiban, I., Yu, C., Maver, K.G., Maver, M.G.: An automated information retrieval
platform for unstructured well data utilizing smart machine learning algorithms
within a hybrid cloud container. In: First EAGE/PESGB Workshop Machine
Learning (2018)
17. Hong, K., Lillethun, D., Ramachandran, U., Ottenwälder, B., Koldehofe, B.:
Opportunistic spatio-temporal event processing for mobile situation awareness. In:
Proceedings of the 7th ACM International Conference on Distributed Event-based
Systems (2013)
18. Mao, J., Tian, W., Zhang, Y., Cui, J., Ma, H., Bian, J., Liu, J., Zhang, J.: Co-
check: collaborative outsourced data auditing in multicloud environment. Secur.
Commun. Netw. 2017, 13 (2017)
19. Modi, K.J., Kapadia, N.: Securing healthcare information over cloud using hybrid
approach. In: Panigrahi, C., Pujari, A., Misra, S., Pati, B., Li, K.C. (eds.) Progress
in Advanced Computing and Intelligent Engineering. Springer, Singapore (2019)
20. Opara-Martins, J.: Taxonomy of cloud lock-in challenges. In: Khatib, M., Salman,
N. (eds.) Mobile Computing. IntechOpen, Rijeka (2018)
21. Opara-Martins, J., Sahandi, R., Tian, F.: Critical analysis of vendor lock-in and its
impact on cloud computing migration: a business perspective. J. Cloud Comput.
5(1), 4 (2016)
22. Ouyang, C., Moura, M.: A vision of hybrid cloud for big data and analytics. Tech-
nical report, IBM Big Data and Analytic Hub (2017)
23. Petcu, D.: Portability and interoperability between clouds: challenges and case
study. In: Towards a Service-Based Internet. Springer, Berlin (2011)
24. Petcu, D.: Multi-cloud: expectations and current approaches. In: Proceedings of
the International Workshop on Multi-cloud Applications and Federated Clouds.
Prague, Czech Republic (2013)
25. Peter, M., Tim, G.: The NIST definition of cloud computing. Technical report,
National Institute of Standards and Technology Gaithersburg, MD 20899-8930
and U.S. Department of Commerce (2011)
26. Qi, Z.: Cloud computing: state-of-the-art and research challenges. J. Internet Serv.
Appl. 1(1), 7–18 (2010)
27. Sun, Y., Zhang, J., Xiong, Y., Zhu, G.: Data security and privacy in cloud com-
puting. Int. J. Distrib. Sens. Netw. 10(7), 190903 (2014)
1068 J. Hong et al.
28. Taherizadeh, S., Jones, A.C., Taylor, I., Zhao, Z., Stankovski, V.: Monitoring self-
adaptive applications within edge computing frameworks: a state-of-the-art review.
J. Syst. Softw. 136, 19–38 (2018)
29. Taherkordi, A., Zahid, F., Verginadis, Y., Horn, G.: Future cloud systems design:
challenges and research directions. IEEE Access 6, 74120–74150 (2018)
30. Thillaiarasu, N., ChenthurPandian, S.: Enforcing security and privacy over multi-
cloud framework using assessment techniques. In: Proceedings of the 10th Interna-
tional Conference on Intelligent Systems and Control (ISCO), India. IEEE, Coim-
batore (2016)
31. Vaquero, L.M., Rodero-Merino, L.: Finding your way in the fog: towards a compre-
hensive definition of fog computing. SIGCOMM Comput. Commun. Rev. 44(5),
27–32 (2014)
32. Varghese, B., Buyya, R.: Next generation cloud computing: new trends and
research direction. Future Gener. Comput. Syst. 79, 849–861 (2018)
33. Vurukonda, N., Rao, B.T.: A study on data storage security issues in cloud com-
puting. Procedia Comput. Sci. 92, 128–135 (2016)
34. Winans, T.B., Brown, J.S.: Cloud Computing: A collection of Working Papers.
Deloitte LLC, New York (2009)
35. Yi, S., Li, C., Li, Q.: A survey of fog computing: concepts, applications and issues.
In: Proceedings of the 2015 Workshop on Mobile Big Data, China. ACM, Hangzhou
(2015)
36. Zhu, Y., Hu, H., Ahn, G.J., Han, Y., Chen, S.: Collaborative integrity verification in
hybrid clouds. In: 7th International Conference on Collaborative Computing: Net-
working, Applications and Worksharing (CollaborateCom), USA. IEEE, Orlando
(2012)
The 1st Workshop on Artificial
Intelligence and Machine Learning
(AIMAL-2019)
Multi-objective Optimal Power Flow
Using Improved Multi-objective
Multi-verse Algorithm
1 Introduction
The optimal power flow (OPF) is considered a vital optimization tool for efficient
operation and planning of the power system. The OPF is used to determine the
control variables at a specific time for the power system operation in order
to optimize (maximize or minimize) different objective functions (OFs) while
assuring the technical feasibility of the solution [1].
Limited reserves of fossil fuels and environmental protection issues have
raised the concern to integrate renewable energy resources (RESs) into the grid.
Wind power plants and solar photovoltaic are amongst the most common RESs.
A lot of research work has been done on classical OPF, which consider only
thermal generators. Classical OPF, which is by itself non-linear and non-convex
and large scale optimization problem because of different OFs and constraints,
however, by incorporating uncertain RESs further increase the complexity of the
problem [2].
The OPF was firstly introduced in 1962 by Carpentier. Since the inception of
OPF, many optimization methods have been introduced to the solution of OPF
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1071–1083, 2019.
https://doi.org/10.1007/978-3-030-15035-8_104
1072 M. Abdullah et al.
problems. These approaches can be divided into classical numerical analysis and
modern metaheuristics algorithms. The classical method includes linear pro-
gramming, quadratic programming, interior point, and dynamic programming
etc. These techniques are inadequate for large-scale problems because they are
comprised of complex, lengthy calculation and they are excessively dependent on
initial guess values. In the past few years, different metaheuristic algorithms have
been developed to overwhelm the deficiencies of the classical methods [2,3]. Some
of these algorithms are; whale optimization algorithm (WOA), particle swarm
optimization algorithm (PSO), moth-swarm algorithm (MSA), modified firefly
algorithm (MODFA), non-dominating sorting genetic algorithm (NSGA), differ-
ential evolution (DE), elephant herding optimization (EHO) technique, Teaching
learning-based optimization (TLBO) technique etc. In any case, due to the incon-
sistency of the OPF solution for different objectives, no heuristic algorithm can
be proven optimal for solving all these objectives, Hence, there is always a room
for new techniques and algorithms that can effectively solve multiple objectives
of OPF.
The OPF problem is considered as single-objective if only one objective func-
tion is reckoned to be optimized. However, OPF problem needs to optimize
conflicting objectives in real world, simultaneously. For solving multi-objective
optimization (MOO) problems, two techniques are used in the literature: a pri-
ori and a posteriori [4]. In the first MOO technique, the MOO problem is con-
verted into a single objective (SO) by assigning weight to different objectives and
aggregating them. The importance of each objective is dictated by the weights
assigned to it, which is generally set by an expert. Solving the problem with this
approach does not require any modification in the algorithm. In posteriori MOO
approach, there is no need to define the weights by an expert. In this method,
different solutions are there for the decision maker as compared to the former
one. There is no need to define weights for each objective by an expert and the
Pareto frontier is approximated, even in a single run [5,6].
Increasing the participation of consumers in the electrical network operations
and accommodation of RESs are the most important features of smart grids. The
role of OPF in smart grids is very vital in order to decrease the transmission,
distribution losses, emission of fossil fuels, and the energy generation cost. Sub-
sequently, decreasing the electrical power consumption and prices, concurrently,
enabling consumers and electrical utilities to regulate the demand [7–11].
subject to:
gk (x́, v́) = 0 k = 1, 2, . . . , G (2)
hl (x́, v́) 0 l = 1, 2, . . . , L, (3)
th
where, fobj are the OFs to be minimized, gk (x́, v́) is the k equality constraint
(EC) which actually represent the load flow equations [14]. The inequality con-
straints (IC) are expressed by h(x́, v́). x́ and v́ are the vectors of state (dependent)
variables and control (independent) variables respectively [15].
The state variables are:
Besides these, tap settings of transformer, and compensation of shunt VAR (reac-
tive power) compensators can also be considered as control parameters.
Accordingly, vector v́ can be expressed as:
Two OFs are considered in this paper, i.e. minimization of generation cost and
power loss, whose details are provided subsequently.
operated in sequence, which results in the non-continuous cost curve. The cost
model of thermal generators considering valve loading effect is given by [17]:
N
TG
where, NT G is the number of thermal generators. ái , b́i , and ći are the cost
coefficient of ith generator, producing output power PT Gi , d´i and éi are the
coefficients of valve point loading effect. The values of thermal units coefficients
are listed in Table 1.
Cost coefficients á b́ ć d́ é
GT h1 0 2 0.00375 18 0.037
GT h2 0 1.75 0.0175 16 0.038
GT h8 0 3.25 0.00834 12 0.045
The third renewable energy resource that is considered in this study is the
combination of wind and run of river small hydro generation plant. The scheduled
power is jointly delivered by wind and hydro generators. The hydro generation
considered in this work is of 5 MW, which is proportional to the river’s flow rate
[20]. The direct cost of the combined wind and hydro plant is described as:
where, PW H represent the scheduled real power from the combined plant, PW H,w
and PW H,h are the electrical energy contribution from the wind and hydro units
respectively. gw is the direct cost coefficient of wind unit and gh is that of hydro
unit.
As wind energy is stochastic in nature, the actual power produced may be less
or more than scheduled power. Therefore, the ISO must have reserve generating
capability to meet the demand. Reserve cost for the wind unit can be calculated
as [21]:
CRW,i (PW sh,i − PW ac,i ) = krw,i (PW sh,i − PW ac,i )
PW sh,i (8)
= krw,i (PW sh,i − pw,i )fw (pw,i )dpw,i .
0
In the case when the output power of the wind generators is greater then the
scheduled power, the ISOs pay penalty cost if the surplus power is not utilized
by them by reducing the power of thermal generators.
The penalty cost of wind plant can be mathematically described as:
CPW,i (PW ac,i − PW sh,i ) = kpw,i (PW ac,i − PW sh,i )
PW r,i
(9)
= kpw,i (pw,i − PW sh,i )fw (pw,i )dpw,i ,
PW sh,i
where, PW sh,i is the scheduled power, PW ac,i is the available power from wind
plant, PW r,i is the rated power, and fw (pw,i )is the probability density function
of wind power. Similarly the reserve and penalty cost of solar and combined
wind and hydro can be calculated. The values of direct, reserve, and penalty
cost coefficients for RESs are provided in Table 2.
The output power of wind power units is directly proportional to wind speed.
Wind speed probability is predominantly represented by Weibull, probability
distribution function (PDF) [17,21,22].
The power given by PV arrays depends on solar irradiance. The probability
distribution of solar irradiance Gs is represented by lognormal PDF [21,23].
Gumbel distribution correctly represent the river flow rate [24,25]. Summary of
the number of turbines, the rated power output from RESs, and the values of
PDF parameters are listed in Table 3.
Where nL is the number of branches, Gij is the conductance of the branch which
linked together the ith and j th bus, and δij is the voltage angle between them.
Vi and Vj are the voltages of bus i and bus j. The magnitude of these voltages
are calculated through newton-Raphson load flow equation.
1076 M. Abdullah et al.
and
NB
QG − QD − Vl Vl [Gij cos(δij + βij sin(δij ))] = 0 ∀i N B. (12)
j=1
PTmin max
G,i PT G,i PT G,i i = 1, . . . , NT G , (13)
min max
PW G,j PW G PW G , (14)
PPmin max
V,k PP V PP V , (15)
min max
PW HG,j PW HG,j PW HG,j , (16)
Qmin max
T G,i QT G,i QT G,i , i = 1, . . . , NT G , (17)
Qmin max
W G QW G QW G , (18)
Qmin max
P V QP V QP V , (19)
Qmin
W HG QW HG Qmax
W HG , (20)
min max
VG,i VG,i VG,i , i = 1, . . . , NG , (21)
where, PT G,i is the active power generation of ith thermal unit, PW G , PP V , and
PW HG is the real power of wind farm, solar PV array, and combined wind and
hydro generators, respectively. Similarly, QT G,i , QW G , QP V , and QW HG are the
reactive power output of the thermal, wind, solar, and combined wind and hydro
generators. VG,i is the voltage of the ith generation bus.
Security constraints includes the voltage limits of P-Q buses and the trans-
mission lines loading capability, which are described as follow:
min max
VLp VLp VLp p = 1, . . . , NLB , (22)
and
max
Slq Slq q = 1, . . . , Nl . (23)
Equation 22 represents the voltage limits for NLB number of load buses and
the line loading constraint is described in Eq. 23 for Nl branches.
1078 M. Abdullah et al.
Table 4. Best results of two-objective for Case 1 of 30 bus power system at 500
iteration.
used to maintain the diversity of the non-dominated solution. The solutions are
truncated to the size of the archive to reduce computation time. The new solu-
tions are compared to the stored results in the archive and the better solutions
are replaced in the archive.
Modified IEEE 30-bus test system is considered to varify the effectiveness of the
proposed IMOMVO for the solution of MOOPF and its performance is compared
with the results of MOPSO, MOMVO, and NSGA-II optimization algorithm.
The codes of these algorithms are written in the MATLAB environment and
MATPOWER (MATLAB package) is used for power flow calculation. All the
simulations are performed on Intel Core i7, 2.00 GHz processor with an 8 GB
RAM personal computer. For the fair comparison of the proposed IMOMVO
with MOPSO, MOMVO, and NSGAII in terms of obtaining the best PF. The
size of the population (Np = 50) and numbers of iterations are taken the same
for all of these algorithms. The summary of the adapted 30-bus is provided in
Table 5.
Two conflicting objectives, total generation cost minimization, and power
loss minimization are solved simultaneously using proposed IMOMVO and its
performance is compared with other optimization algorithms. The PF obtained
from the simulation of these algorithms for solving MOOPF problem of modified
IEEE 30-bus system is shown in Figs. 1, 2, and 3 for 20, 100, and 500 runs. From
the simulation results, it is obvious that the PF obtained from IMOMVO per-
formed better than MOPSO, MOMVO, and NSGA-II, both at lower and higher
iteration (times) in term of solution optimality and distribution. The values of
the control variables for the best cost, best power loss, and the resultant OFs are
provided in Table 4. The PF obtained by MOMVO is better than MOPSO and
NSGA-II at lower iteration, however, at higher iteration, the PF obtained by
MOMVO is not well distributed. The proposed IMOMVO eliminated the limi-
tation of PF distribution of MOMVO algorithm by incorporating the crowding
distance and non-dominating sorting approaches and displayed the best results
both at lower and higher iteration.
4.1 Conclusion
In this study, IMOMVO algorithm has been proposed and applied to solve
MOOPF problem. The proposed algorithm was successfully implemented on
modified IEEE 30-bus power systems. The simulation results reveal the superi-
ority of the IMOMVO over NSGA-II, MOPSO and MOMVO algorithms. The
IMOMVO eliminated the limitation of PF distribution of MOMVO algorithm
and displayed the better results both at lower and higher iteration. Therefore,
the IMOMVO provides better PF. The PF obtained by multi-objective algo-
rithms helps the decision maker to take a better informed-decision, concerning
the compromise between the conflicting objectives.
1082 M. Abdullah et al.
References
1. Abdi, H., Beigvand, S.D., La Scala, M.: A review of optimal power flow studies
applied to smart grids and microgrids. Renew. Sustain. Energy Rev. 71, 742–766
(2017)
2. Attia, A.-F., El Sehiemy, R.A., Hasanien, H.M.: Optimal power flow solution in
power systems using a novel Sine-Cosine algorithm. Int. J. Electr. Power Energy
Syst. 99, 331–343 (2018)
3. Bai, W., Eke, I., Lee, K.Y.: An improved artificial bee colony optimization algo-
rithm based on orthogonal learning for optimal power flow problem. Control Eng.
Pract. 61, 163–172 (2017)
4. Branke, J., Kaußler, T., Schmeck, H.: Guidance in evolutionary multi-objective
optimization. Adv. Eng. Softw. 32(6), 499–507 (2001)
5. Tan, K.C.: Advances in evolutionary multi-objective optimization. In: Soft Com-
puting Applications. Springer, Heidelberg, pp. 7–8 (2013)
6. Mirjalili, S., et al.: Salp swarm algorithm: a bio-inspired optimizer for engineering
design problems. Adv. Eng. Softw. 114, 163–191 (2017)
7. Javaid, N., et al.: Energy efficient integration of renewable energy sources in the
smart grid for demand side management. IEEE Access 6, 77077–77096 (2018)
8. Javaid, N., et al.: An intelligent load management system with renewable energy
integration for smart homes. IEEE Access 5, 13587–13600 (2017)
9. Khan, M., et al.: Game theoretical demand response management and short-term
load forecasting by knowledge based systems on the basis of priority index. Elec-
tronics 7(12), 431 (2018)
10. Awais, M., et al.: Towards effective and efficient energy management of single home
and a smart community exploiting heuristic optimization algorithms with critical
peak and real-time pricing tariffs in smart grids. Energies 11(11), 3125 (2018)
11. Ahmad, A., et al.: An optimized home energy management system with integrated
renewable energy and storage resources. Energies 10(4), 549 (2017)
12. Nuaekaew, K., et al.: Optimal reactive power dispatch problem using a two-archive
multi-objective grey wolf optimizer. Expert Syst. Appl. 87, 79–89 (2017)
13. Kumar, A.R., Premalatha, L.: Optimal power flow for a deregulated power system
using adaptive real coded biogeography-based optimization. Int. J. Electr. Power
Energy Syst. 73, 393–399 (2015)
14. Chen, G., et al.: Quasi-oppositional cuckoo search algorithm for multi-objective
optimal power flow. IAENG Int. J. Comput. Sci. 45(2), 255–266 (2018)
15. Ghasemi, M., et al.: Multi-objective optimal power flow considering the cost, emis-
sion, voltage deviation and power losses using multi-objective modified imperialist
competitive algorithm. Energy 78, 276–289 (2014)
16. Abdelaziz, A.Y., Ali, E.S., Abd Elazim, S.M.: Flower pollination algorithm to
solve combined economic and emission dispatch problems. Eng. Sci. Technol. Int.
J. 19(2), 980–990 (2016)
17. Biswas, P.P., Suganthan, P.N., Amaratunga, G.A.J.: Optimal power flow solutions
incorporating stochastic wind and solar power. Energy Convers. Manag. 148, 1194–
1207 (2017)
18. Chen, C.-L., Lee, T.-Y., Jan, R.-M.: Optimal wind-thermal coordination dispatch
in isolated power systems with large integration of wind capacity. Energy Convers.
Manag. 47(18–19), 3456–3472 (2006)
19. Wijesinghe, A., Lai, L.L.: Small hydro power plant analysis and development. In:
2011 4th International Conference on Electric Utility Deregulation and Restruc-
turing and Power Technologies (DRPT). IEEE (2011)
Multi-objective Optimal Power Flow 1083
20. Tanabe, R., Fukunaga, A.: Success-history based parameter adaptation for differ-
ential evolution. In: 2013 IEEE Congress on Evolutionary Computation (CEC).
IEEE (2013)
21. Reddy, S.S., Bijwe, P.R., Abhyankar, A.R.: Real-time economic dispatch consider-
ing renewable power generation variability and uncertainty over scheduling period.
IEEE Syst. J. 9(4), 1440–1451 (2015)
22. Reddy, S.S.: Optimal scheduling of thermal-wind-solar power system with storage.
Renew. Energy 101, 1357–1368 (2017)
23. Chang, T.P.: Investigation on frequency distribution of global radiation using dif-
ferent probability density functions. Int. J. Appl. Sci. Eng. 8(2), 99–107 (2010)
24. Mujere, N.: Flood frequency analysis using the Gumbel distribution. Int. J. Com-
put. Sci. Eng. 3(7), 2774–2778 (2011)
25. Cabus, P.: River flow prediction through rainfall runoff modelling with a
probability-distributed model (PDM) in Flanders, Belgium. Agric. Water Manag.
95(7), 859–868 (2008)
26. Ghasemi, M., et al.: Solving non-linear, non-smooth and non-convex optimal power
flow problems using chaotic invasive weed optimization algorithms based on chaos.
Energy 73, 340–353 (2014)
27. Mohamed, A.A.A., et al.: Optimal power flow using moth swarm algorithm. Electr.
Power Syst. Res. 142, 190–206 (2017)
28. Mirjalili, S., et al.: Optimization of problems with multiple objectives using the
multi-verse optimization algorithm. Knowl.-Based Syst. 134, 50–71 (2017)
29. Mirjalili, S., Mirjalili, S.M., Hatamlou, A.: Multi-verse optimizer: a nature-inspired
algorithm for global optimization. Neural Comput. Appl. 27(2), 495–513 (2016)
30. Yang, X.-S.: Flower pollination algorithm for global optimization. In: International
Conference on Unconventional Computing and Natural Computation. Springer,
Heidelberg (2012)
Enhanced Robustness Strategy for IoT
in Smart Cities Based on Data
Driven Approach
1 Introduction
Revolutionary advances in telecommunication, information technology, electron-
ics and applications resulted into formation of Internet. Initially it was built
with an aim to connect people. However by 2008, it was found that it has
connected even more devices in comparison to people. Exponential growth of
concept primarily been as Internet of Things (IoT). 50 billion new connections
are predicted to be made till 2020 by IoT [1]. IoT [2,3] is a merger of numer-
ous concepts including fifth generation cellular network, hybrid wireless mobile
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1084–1096, 2019.
https://doi.org/10.1007/978-3-030-15035-8_105
Enhanced Robustness Strategy 1085
A. By using data driven approach from big data servers of smart cities, scheme
for enhancing robustness of scale-free IoT networks is proposed.
B. Proposed enhancement in ROSE [15] will enhance the robustness of scale-free
network for IoT without changing node degree distribution. Proposed EASO
ROSE scheme further enhances the topology towards onion-like structure.
Algorithm proven to be more robust against malicious attacks against all
other schemes including ROSE and Simulating Annealing (SA).
Rest of the paper is arranged as follows. Section 2 is related work in which past
efforts in the same domain is presented. Section 3 explains the applied modeling
strategy of IoT network. Topology robustness matrices are discussed in Sect. 4.
Details of our first topology robustness enhancement scheme is presented as
EASO ROSE in Sect. 5. Evaluation results are shown in Sect. 6. Finally conclu-
sion is extracted in Sect. 7 (Fig. 1).
1086 T. N. Qureshi et al.
2 Related Work
Barabasi and Albert proposed a model named as Barabasi-Albert (BA) model
[17]. To generate power-law distribution of node degree in scale-free networks,
BA model follows the following criteria to achieve scale-free characteristics.
• The nodes are added one by one to the network, at each successive step a
new node joins the network such that m ≤ No . m is number of existing nodes
and No total number of nodes.
• When new node joins the existing network, it chooses node to connect based
on probability that a new node will connect to i node depends on its degree ki .
This means that node having more connections are more likely to be chosen
as next hop for new joining node. This is termed as preferential attachment
and concept is termed as M athewEf f ect.
Small number of nodes are having high degree in scale-free networks. Mali-
cious and targeted attacks are intended to destroy high degree nodes. Failing
high degree node results in removal of large number of edges at the same time.
This leads to fragmentation in the network. Issue can be catered by adding
additional edges between nodes. However, it require additional energy resources,
wireless channels and bandwidth. Hence increases cost and additional edges also
destroy scale-free property of the network. Therefore enhancing topology robust-
ness against malicious attacks requires that initial degree distribution of network
remains unchanged after optimization. Multiple modeling strategies to enhance
the robustness of topology have been proposed by many researchers in recent
years.
Schneider et al. [18] proposed a new metric R to measure robustness of
a network. Largest connected cluster is considered after each attack. Metric
R is defined as number of largest connected nodes to the proportion of total
Enhanced Robustness Strategy 1087
Where Degree(i) is degree of ith node and n is the total number of live
nodes in the network. When a new node joins the network, it gets the neighbor
information through the big data servers in the smart city after obtaining its
own location through Global Positioning System (GPS) circuit. All the sensor
broadcast its location after specific interval of time to big data servers. Big
data servers contains the geographic locations of all the sensors of multiple IoT
networks in smart city.
In the above equation N represents the total number of nodes and s(Q) is
the fraction of largest connecting cluster nodes or nodes with largest connected
subgraphs, means highest degree. 1/N is the normalization factor that ensures
that networks of different sizes can be compared. Range of R for any network
can be varied between 1/N to 0.5 numerical value. Where 1/N limit represents
the start topology and 0.5 represents the full mesh topology with fully connected
graph.
Fig. 2. Swapping possibilities of independent edges, (a) Original topology, (b) Swap-
ping possibility 1, (c) Swapping possibility 2
In onion-like structure nodes are having edges forming the ring formation. In
addition to the ring formation, there exist some edges that are horizontal with
respect to core of onion, ROSE [15] put these horizontal edges into considera-
tion. Through enhanced angle sum operation, topology is converged to onion-
like structure by using these horizontal edges. To begin with the enhanced angle
sum operation EASO ROSE selects centroid node and take centroid as reference.
According to the Schneider et al. [18] and Zhou et al. [27], onion-like structure
consists of core of highest degree nodes surrounded by rings of low degree nodes
as we move towards outer side of the onion. Through simulation and detailed
analysis, it is found that many times first nodes may not belong to core of highest
degree nodes. To cater for the issue, EASO ROSE fetches the degree list of all
nodes in IoT from big data servers of smart cities. Then calculates the highest
degree nodes. High degree node is calculated by following formula after fetching
node degree list.
Highesti = max(Listnode−degree ) (7)
Then highest degree node is selected as centroid is given by:
β = 180o − α (10)
By the same process as for edge eij , we calculate the both angles α and β for
the other edge ekl as shown in Fig. 3(c). For all the possible swap combination
of edges between nodes i, j, k, l we calculate the both angles α and β for both
edges as per Fig. 3(d). The combination with largest α is accepted for possible
swap. For assurance Schneider R is again calculated.
1092 T. N. Qureshi et al.
Fig. 3. (a) Geometric details of angle sum operation (b) First edge swap possibility,
(c) Second edge swap possibility, (d) Third edge swap possibility
0.13
0.12
Schneider R
0.11
EASO-ROSE
ROSE
0.1
Simulating Annealing
0.09
0.08
0.07
0.06
0 10 20 30 40 50 60 70 80 90 100
Number of Iterations
0.25
EASO-ROSE
ROSE
0.2 Simulating Annealing
Schneider R
0.15
0.1
0.05
Number of Nodes
Fig. 6. Comparison between EASO ROSE and other algorithms with different sizes of
network topologies in IoT
1094 T. N. Qureshi et al.
7 Conclusion
Considering the practical requirements of smart cities for IoT network, we
present our study. Based on data driven concept from big data servers, we extract
node geographic information and neighbor list. Then initial IoT topology is con-
structed. Enhancements in ROSE is suggested and validated by simulations.
Enhancement proved to improve robustness of IoT network topology.
We have designed enhanced criteria of angle sum operation phase of ROSE.
Initial degrees are unchanged for all nodes during the operation. Hence additional
energy cost of adding edges are saved. The enhancement further converts the
topology into onion like structure and increases robustness against malicious
attacks.
Finally the performance of EASO ROSE is tested over scale-free network
typologies of different sizes and edge densities. Performance is compared with
two algorithms including classic ROSE and simulating annealing. Simulation
results proves that our proposed algorithm enhances the topology robustness of
scale-free networks without disturbing the degree distribution.
References
1. Collier, S.E.: The emerging enernet: convergence of the smart grid with the internet
of things. In: 2015 IEEE Rural Electric Power Conference (REPC) (2015)
2. Stankovic, J.A.: Research directions for the Internet of Things. IEEE Internet
Things J. 1(1), 3–9 (2014)
3. Javaid, N., Sher, A., Nasir, H., Guizani, N.: Intelligence in IoT based 5G networks:
opportunities and challenges. IEEE Commun. Mag. 56(10), 94–100 (2018)
4. Naz, A., Javaid, N., Iqbal, M.M., Ali, M., Imran, M., Khan, Z.A.: TBEENISH:
threshold balanced energy efficient network integrated super heterogeneous proto-
col for WSNs. In: 32nd International Conference on Advanced Information Net-
working and Applications (IEEE AINA) (2018)
5. Butt, S.A., Bakar, K.A., Javaid, N., Gharaei, N., Ishmanov, F.H., Afzal, M.K.,
Mahmood, M.K., Mujahid, M.A.: Exploiting layered multi-path routing protocols
to avoid void hole regions for reliable data delivery and efficient energy management
for IoT-enabled underwater WSNs. Sensors 19, 510 (2019)
6. Mateen, A., Awais, M., Javaid, N., Ishmanov, F., Afzal, M.K., Kazmi, S.:
Geographic and opportunistic recovery with depth and power transmission adjust-
ment for energy-efficiency and void hole alleviation in UWSNs. Sensors 19(3), 709
(2019)
7. Hussain, R., Bouk, S.H., Javaid, N., Khan, A., Lee, J.: Realization of VANET-
based clouds services through named data networking. IEEE Commun. Mag. 58,
168–175 (2018)
Enhanced Robustness Strategy 1095
8. Sher, A., Khan, A., Javaid, N., Ahmed, S.H., Aalsalem, M., Khan, W.Z.: Void
hole avoidance for reliable data delivery in IoT enabled underwater wireless sensor
networks. Sensors 18, 3271 (2018)
9. Qureshi, T.N., Javaid, N.: Enhanced adaptive geographic opportunistic routing
with interference avoidance assisted with mobile sinks for underwater wireless sen-
sor network. In: International Conference on Frontiers of Information Technology
(2018)
10. Qureshi, T.N., Javaid, N., Khan, A.H., Iqbal, A., Akhtar, E., Ishfaq, M.: Balanced
energy efficient network integrated super heterogenous protocol for wireless sensor
networks. In: International Workshop on Body Area Sensor Networks (BASNet)
(2013)
11. Naz, A., Javaid, N., Qureshi, T.N., Imran, M., Ali, M., Khan, Z.A.: EDHBPSO:
enhanced differential harmony binary particle swarm optimization for demand side
management in smart grid. In: 32nd International Conference on Advanced Infor-
mation Networking and Applications (IEEE AINA) (2018)
12. Qureshi, T.N., Javaid, N., Naz, A., Ahmad, W., Imran, M.: A novel meta-heuristic
hybrid enhanced differential harmony wind driven (EDHWDO) optimization tech-
nique for demand side management in smart grid. In: 32nd International Confer-
ence on Advanced Information Networking and Applications (IEEE AINA) (2018)
13. Qiu, T., Liu, J., Si, W., Han, M., Ning, H., Atiquzzaman, M.: A data-driven robust-
ness algorithm for the internet of things in smart cities. IEEE Commun. Mag.
(2017)
14. Wu, J., et al.: Spectral measure of structural robustness in complex networks. IEEE
Trans. Syst. Man Cybern. Part A Syst. Hum. 41(6), 1244–1252 (2011)
15. Qiu, T., Zhao, A., Xia, F., Si, W., Do, W.: ROSE: robustness strategy for scale-free
wireless sensor networks. IEEE/ACM Trans. Networking 55, 18–23 (2017)
16. Li, R.H., Yu, J.X., Huang, X., Cheng, H., Shang, Z.: Measuring robustness of com-
plex networks under MVC attack. In: Proceedings of the 21st ACM International
Conference on Information and Knowledge Management, New York, NY, USA, pp.
1512–1516, October/November 2012
17. Barabási, A.-L., Albert, R.: Emergence of scaling in random networks. Science
286(5439), 509–512 (1999)
18. Schneider, C.M., Moreira, A.A., Andrade Jr., J.S., Havlin, S., Herrmann, H.J.:
Mitigation of malicious attacks on networks. Proc. Nat. Acad. Sci. USA 108(10),
3838–3841 (2011)
19. Buesser, P., Daolio, F., Tomassini, M.: Optimizing the robustness of scale-free
networks with simulated annealing. In: Proceedings of the 10th International Con-
ference on Adaptive and Natural Computing Algorithms (ICANNGA), Ljubljana,
Slovenia, pp. 167–176, April 2011
20. Liu, L., Qi, X., Xue, J., Xie, M.: A topology construct and control model with small-
world and scale-free concepts for heterogeneous sensor networks. Int. J. Distrib.
Sens. Netw. 10(3), 1–8 (2014). Art. no. 374251
21. Xiao, W., Lin, L., Chen, G.: Vertex-degree sequences in complex networks: new
characteristics and applications. Physica A 437(11), 437–441 (2015)
22. Frank, H., Frisch, I.T.: Analysis and design of survivable networks. IEEE Trans.
Commun. Tech. 18, 501–519 (1970)
23. Latora, V., Marchiori, M.: Efficient behavior of small-world networks. Phys. Rev.
Lett. 87, 198701 (2005)
24. Sydney, A., Scoglio, C., Youssef, M., Schumm, P.: Characterizing the robustness
of complex networks. IJITST 2, 291–320 (2010)
1096 T. N. Qureshi et al.
25. Fiedler, M.: Algebraic connectivity of graphs. Czech Math. J. 23, 298–305 (1973)
26. Holme, P., Kim, B.J., Yoon, C.N., Han, S.K.: Attack vulnerability of complex
networks. Phys. Rev. E 65, 056109–056123 (2002)
27. Zhou, M., Liu, J.: A memetic algorithm for enhancing the robustness of scale-free
networks against malicious attacks. Phys. A Statist. Mech. Appl. 410, 131–143
(2014)
28. Prendeville, S., Cherim, E., Bocken, N.: Circular cities: mapping six cities in tran-
sition. Environ. Innov. Societal Transitions 26, 171–194 (2018)
Game-Theoretical Energy Management
for Residential User and Micro Grid
for Optimum Sizing of Photo Voltaic
Battery Systems and Energy Prices
1 Introduction
In order to achieve quality, traditional grid is transforming into smart grid.
Smart grid is considered as cyber physical system that contain different phys-
ical system such as energy production, energy distribution and energy usage.
Smart grid also contain several advanced techniques such as advanced metering
infrastructure (AMI), energy management system as well as electrical vehicle
[1]. Smart grid is different from traditional grid in terms of functionality and
it works centrally in bidirectional manner. Customer actively involves in distri-
bution of energy by announcing total energy requirement. Large electrical grid
is divided into small scale grid in smaller geographical areas [2,3,19,20]. Small
scale grids are also known as microgrid. Microgrid generates energy with the
help of renewable resources such as wind energy, hydro power and solar power.
That is the reason that microgrid total amount of energy is not defined as it
depends on the weather and climatic conditions. There can be the case when
microgrid does not produce sufficient energy and it purchase energy from utility
in order to full fill requirement of the users [4]. At the same time, if microgrid
is having surplus energy it will sale it to utility. In return, utility will award it
with certain subsidy. In real world energy management system contain different
kind of errors include, implementation, estimation errors. Such kind of errors
make system completely meaningless. Therefore, it is very important to bring
forecasting of energy generation in account along with its expected error. It helps
in performing contingency planning for better energy management [5,21,22]. In
our proposed work, game theoretic energy management system is proposed. It
contains two level players that includes user and micro grid. Back Propagation
(BP), Support Vector Machine (SVM) and Stacked Auto Encoder (SAE) fore-
casting techniques are used to forecast future solar power generation.
2 Contribution
This paper has following contribution: 1. It considers two stage stackelberg game
in which microgrid is leader and user is follower. Besides, MG may purchase
electricity from utility. If load demand of user exceeds its generation capacity as
shown in Figs. 1 and 2. Two scenarios are considered in which energy manage-
ment with and without GT are discussed. In addition energy storage mechanism
is employed to reduce energy cost of MG. Game theory optimize payoff of both
player by the help of predicted energy generated by Photo voltaic (PV) cell.
power for residential user and charging the battery. It also send surplus energy
to the utility. After receiving utility price policy from information network, user
sends demand to micro grid and eventually that demand will be sent to utility.
In proposed system model. User in this scenario have shiftable and unshiftable
loads [17,18]. One day, as a period is taken for scheduling energy consumption
at user end. K denotes each time slots in a period. Here, the utility accepts
electricity demand from user for each time slots and real time price are sent to
user against each time slot. {P k = pk1 , ...pkj , ....pkm }. Microgrid set its prices to
maximize the pay off as per demand of each user in real time.
is cheaper than the power that is supplied by utility company. Priorities for
providing solar power at any working time in day light to residential users N
by microgrid is: Foremost priority is to be given to fulfill demand of N users
and to charge the batteries, at second priority, surplus energy is to be sold to
utility for generating revenue by trading. For microgrid, lets assume solar power
generation is en (k) ≥ 0 in time slot k. Solar power generation provides energy
en (k) − eln (k) − ebn (k) = 0 means microgrid does not have surplus energy to
be sold to utility. In case en (k) − eln (k) − ebn (k) ≥ 0 means microgrid is having
surplus power that is generated by solar panel is to be sold. Profit of microgrid
that it generates while selling surplus power to utility is:
K
Un = λs en (k) − eln (k) − ebn (k)) (3)
k=1
Microgrid
Solar Power
Solar Panel for
Generation Electricity
forcasting using Generation
historical data
D(x,k)
Battery Storage for
Solar Generated
Solar generated
Power
Power
Residential
Users P(K)
bat
where Cm (y) cost of battery depreciation is represented in form of cents/kWh
and it is also correlated with the material and type of the battery. Battery
depreciation cost is linear increasing function according to the total capacity y
of battery. Apart of battery capacity parameter, there are certain parameters
that are required to be taken under consideration, i.e., charging and discharging
efficiency of battery. Let assume, 0 < ηch < 1 and 0 < ηdisch < 1 shows bat-
tery charging and discharging efficiency. s = [s1 , ..., sk , ..., sK ] represents state of
battery for whole day. Here, battery capacity is also defined therefore inequality
constraint regarding state of battery and capacity of battery is as shown:
0 ≤ sk ≤ z (7)
hkch and hkdisch are binary variable that represents pattern of battery charging
and discharging in each time slot. Battery can either be charged or discharged
at the same time that is shown as:
Proposition 2: For ∀n N and time slot k K, the NE of the non cooperative
game exist and it is also unique.
Proof: As per proof mentioned in [storage game theory, Th. 6], owing to the fact
that cost function Fn (xn , x−n , pk ) is convex in xn , the NE of the non cooperative
game exist and it is also unique.
As shown in Fig. 1, system model is shown in detail. Where micro grid
announces its prices on the basis of solar power generation forecasting. In this
paper, forecasting is performed using BP, SVM and SAE in order to deter-
mine forecasting error. It helps in efficiently forecasting energy generation. Thus
energy management will be more affective.
shows solar power generation in single day. It clearly shows the rise and fall
of the generation as per the solar radiations available in different time slots of
the day. Figure 3 shows the trend of user energy consumption which helps in
deciding the peak hours, off peak hours and mid peak hours respectively. N
users are considered in the scenario which contain both shift able and fixed load
and power consumption of users are shown in Fig. 3. Each day is divided into
24 h time slots k. Consumption of electricity usage at each slot varies depending
upon peak hours and off peak hours.
An analytical scenario is presented to verify the performance and effectiveness
of proposed game theory among user and micro grid. Where N user increases
its pay off by adjusting energy consumption. Besides, micro grid optimizes its
payoff by optimally finalizing energy rates pkm and size of the battery y.
In this paper, TOU pricing scheme is adapted and single day is divided in 24
time slots k. Further, entire day is divided into 3 chunks. Single day is divided
into chunks on the basis of energy consumption, i.e., peak hours, off peak hours
and mid peak hours. Each chunk contain different pricing parameters.
DSM is not discussed in this paper on individual basis. Energy consumption
by user N is catered collectively. However, energy management at micro grid
is discussed in detail that includes energy prices pkm and size of the battery
y. Micro grid is equipped with renewable resources, i.e., photo voltaic cells. In
current scenario, it is assumed that micro grid consist of photo voltaic power
cell in terms of renewable resources.
Microgrid can take more benefit solar power when there are more day light
hours. It also affect battery capacity optimization once microgrid ensures energy
supply to user. However, day light timings and intensity will not be affect the pro-
posed game theoretic energy management. Microgrid employs 2 MW solar power
generation. Energy output by solar cells per day is shown in Fig. 2. Solar power is
Game-Theoretical Energy Management for Residential User 1105
generated during 0600 to 1700 h. User cost relies only on discharging time and it
counts minimal except when battery discharges at peak demand hours. It shows
battery discharging time is from 17:00 to 22:00 h. Charging and discharging of
battery, in both cases in single hour is 1.5 kWh [9], whereas efficiency of battery
charging and discharging is taken as 7.2 cents/kWh [15]. Figure 6 shows mape
values of 3 different forecasting algorithms including BP, SVM, SAE. Historical
data has been used for forecasting solar power that is called step 1, after retrieval
of foretasted data, it is added in historic data and used for prediction that is
termed as step 2 and so on. It is clearly shown in Fig. 6 that mape increases
after each iteration. Simulation shows mape acquired as a result of forecasting
techniques implementation. SAE has performed best among all, therefore error
of SAE is taken as prediction error in objective function of microgrid. There
exist two scenario, that prediction error can be positive, i.e., Δ > 0 and it can
be negative, i.e., Δ < 0 respectively. Δ > 0 shows that predicted output of solar
power is less than the actual. Which leads microgrid to purchase it from utility.
In the second scenario, Δ < 0 is predicted value is greater than actual value.
Which shows microgrid will not procure electricity from utility [12].
5 Conclusion
In this paper, energy management system is focused mainly, which includes
single microgrid and multiple users. To use renewable resources generated energy
optimally, solar power generation forecasting is performed in which three existing
techniques are used and comparison among them is performed. The forecasting
error is, thus brought into account for analyzing the impact on pay off of both
stakeholders in game theory. Stackleberg game theory is utilized in this paper
to prove nash equilibrium among two player, i.e., user and microgrid. In future
work, cooperative energy management will be emphasized.
References
1. Erol-Kantarci, M., Mouftah, H.T.: Smart grid forensic science: applications, chal-
lenges, and open issues. IEEE Commun. Mag. 51(1), 68–74 (2013)
2. Saad, W., Han, Z., Poor, H.V., Basar, T.: Game-theoretic methods for the smart
grid: an overview of microgrid systems, demand-side management, and smart grid
communications. IEEE Sig. Process. Mag. 29(5), 86–105 (2012)
3. Ipakchi, A., Albuyeh, F.: Grid of the future. IEEE Power Energ. Mag. 7(2), 52–62
(2009)
4. Liang, X., Li, X., Lu, R., Lin, X., Shen, X.: UDP: usage-based dynamic pricing
with privacy preservation for smart grid. IEEE Trans. Smart Grid 4(1), 141–150
(2013)
5. Sinha, A., Malo, P., Frantsev, A., Deb, K.: Finding optimal strategies in a multi-
period multi-leader follower Stackelberg game using an evolutionary algorithm.
Comput. Oper. Res. 41, 374–385 (2014)
6. Bingtuan, G.A.O., Xiaofeng, L.I.U., Cheng, W.U., Yi, T.A.N.G.: Game-theoretic
energy management with storage capacity optimization in the smart grids. J. Mod.
Power Syst. Clean Energ. 1–12 (2018)
1106 A. Naz et al.
7. Kanchev, H., Lu, D., Colas, F.: Energy management and operational planning of
a microgrid with a PV-based active generator for smart grid applications. IEEE
Trans. Industr. Electron. 58(10), 4583–4592 (2011)
8. Tazvinga, H., Xia, X.H.: Minimum cost solution of photovoltaic-diesel-battery
hybrid power systems for remote consumers. Sol Energ. 96, 292–299 (2013)
9. http://www.elia.be/en/grid-data/power-generation/Solar-power-generation-
data/Graph
10. Darivianakis, G., Georghiou, A., Smith, R.S., Lygeros, J.: A stochastic optimiza-
tion approach to cooperative building energy management via an energy hub. In:
2015 IEEE 54th Annual Conference on Decision and Control (CDC), pp. 7814–
7819. IEEE, December 2015
11. Valencia, F., Collado, J., Sáez, D., Marı́n, L.G.: Robust energy management system
for a microgrid based on a fuzzy prediction interval model. IEEE Trans. Smart Grid
7(3), 1486–1494 (2016)
12. Zhou, Z., Wang, Y., Wu, Q.J., Yang, C.N., Sun, X.: Effective and efficient global
context verification for image copy detection. IEEE Trans. Inf. Forensic Secur.
12(1), 48–63 (2017)
13. Zhou, Z., Xiong, F., Huang, B., Xu, C., Jiao, R., Liao, B., Yin, Z., Li, J.: Game-
theoretical energy management for energy internet with big data-based renewable
power forecasting. IEEE Access 5, 5731–5746 (2017)
14. Khalid, R., Javaid, N., Rahim, M.H., Aslam, S., Sher, A.: Fuzzy energy manage-
ment controller and scheduler for smart homes. Sustain. Comput. Inf. Syst. 21,
103–118 (2019)
15. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-
ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)
16. Naz, M., Iqbal, Z., Javaid, N., Khan, Z.A., Abdul, W., Almogren, A., Alamri,
A.: Efficient power scheduling in smart homes using hybrid Grey Wolf differential
evolution optimization technique with real time and critical peak pricing schemes.
Energies 11(2), 384 (2018)
17. Nadeem, Z., Javaid, N., Malik, A.W., Iqbal, S.: Scheduling appliances with GA,
TLBO, FA, OSR and their hybrids using chance constrained optimization for smart
homes. Energies 11(4), 888 (2018)
18. Sher, A., Khan, A., Javaid, N., Ahmed, S., Aalsalem, M., Khan, W.: Void hole
avoidance for reliable data delivery in IoT enabled underwater wireless sensor net-
works. Sensors 18(10), 3271 (2018)
19. Mossoud, S., Wollenberg, B.: Toward a smart grid: power delivery for the 21st
century. IEEE Power Energ. Mag. 3, 34–41 (2005)
20. Gungor, V.C., Lu, B., Hancke, G.P.: Opportunities and challenges of wireless sensor
networks in smart grid. IEEE Trans. Ind. Electron. 57(10), 3557–3564 (2010)
21. Wang, Y., Mao, S., Nelms, R.M.: Distributed online algorithm for optimal real-
time energy distribution in the smart grid. IEEE Internet Things J. 1(1), 70–80
(2014)
22. Muralitharan, K., Sakthivel, R., Shi, Y.: Multiobjective optimization technique for
demand side management with load balancing approach in smart grid. Neurocom-
puting 177, 110–119 (2016)
23. Yin, S., Gao, X., Karimi, H.R., Zhu, X.: Study on support vector machine-based
fault detection in Tennessee Eastman process. Abstr. Appl. Anal. 2014, 8 (2014)
24. Dai, Y.M., Gao, Y.: Real-time pricing strategy with multi-retailers based on
demand-side management for the smart grid. Proc. Chin. Soc. Electric. Eng.
34(25), 4244–4249 (2014)
Electricity Load Forecasting for Each Day
of Week Using Deep CNN
1 Background
Forecasting the accurate electricity load demand plays a key role in demand
side management. It helps the generation companies to develop a tolerable pro-
duction plan and avoid energy wastage. For electricity load forecasting many
techniques have been used in literature. These prediction techniques are mainly
classified into three main groups, i.e., data driven techniques, classical techniques
and Artificial Intelligence (AI) techniques. Data driven techniques predict the
desired outcomes on the basis of analysing historical data. Classical techniques
comprises of the statistical and mathematical methods like Autoregressive Inte-
grated Moving Average (ARIMA), Seasonal ARIMA and random forest etc.
Such methods are effective for one step forecast. AI methods inhibit the behavior
of biological neurons e.g., Feed Forward Neural Network (FFNN), Convolution
Neural Network (CNN) and Long Short Term Memory (LSTM) etc.
In [2], Bayesian network and Multi Layer Perceptron (MLP) are used for
short and long term load forecasting. Bayesian network performs behavioral
analytics using the data obtained from the smart meters. Number of experi-
ments were performed on the obtained data. For Short Term Load Forecasting
(STLF), multiple linear regression is used in [3]. However, it has the limitation
that it can not be used for long term prediction. The authors in [4] performed
electricity load forecasting on the basis of weather data obtained from the mete-
orological department. In this work, load prediction is performed using residual
Neural Network (NN). A framework of different models is used for prediction
in [5]. This framework consists of Gated Recurrent Unit (GRU), Recurrent Neu-
ral Network (RNN) and LSTM. GRU is a variant of the RNN. It outperformed
LSTM and many other statistical models in terms of prediction accuracy. Based
on the simulation results, it is proved that deep learning models are better than
the statistical models. The authors in [6] used Restricted Boltzmann machine
and Rectified Linear Unit (ReLU) for training the data and electricity load
prediction respectively. Discrete wavelet transform and inconsistency rate meth-
ods are used for optimal feature selection from the features set in [7]. It helps
in dimensionality reduction. The parameter tuning of Support Vector Machine
(SVM) is performed using Sperm whale algorithm. Authors in [8] proposed a
model for STLF. In this work, for feature selection, Mutual Information (MI) is
used whereas, better prediction results are achieved by modifying the Artificial
Neural Network (ANN). Extreme Learning Machine (ELM) is used in [9] for
efficient load prediction.
The authors in [10] proposed an intelligent approach to forecast electricity
load using MI and ANN. In [11] the authors forecast electricity load and price
using enhanced CNN and enhanced SVM. ELM is a generalized single hidden
layer FFNN learning algorithm. It is proved to be effective for regression and
classification purposes. Authors used the Neural Network (NN) for achieving
better load prediction in [12]. Input weights and biases are randomly assigned
in ELM learning process, whereas output weights are calculated using Moore-
Penrose generalized inverse technique. Efficient load forecasting is performed
using sparse Bayesian ELM in [13]. This method is used for multi-classification
purpose. Authors in [14] used Particle Swarm Optimization (PSO) and discrete
PSO techniques for predicting day ahead electricity load. The Grey wolf opti-
mizer is used in combination with NN to optimize weights and biases in [15].
The simulation results proved that weights and biases optimization technique
improved the prediction accuracy. Back propagation technique is used for train-
ing ELM in [16]. Training is done using context neurons as input to hidden and
input layer of ELM. Authors in [17] enhanced RNN to achieve better prediction
accuracy by adjusting weights using previous iteration errors. Moreover, adjust-
ing biases and neurons selection also affects the prediction accuracy. The authors
in [18] developed a knowledge based system for STLF. In [19], 24 h ahead predic-
tion of cooling load of building is performed using deep learning. The simulation
Electricity Load Forecasting for Each Day of Week Using Deep CNN 1109
results showed that deep learning techniques improved the cooling load predic-
tion. Similarly in [20], RNN is used for electricity load prediction. In this work,
RNN groups the consumers into pool of inputs. It then handles the consumers
data as aggregated data, which reduces the computational time and resource
utilization.
1.2 Contribution
The main contribution of this paper are following:
• An DCNN is proposed to forecast the electricity load of Victoria.
• A performance comparison of the benchmark schemes with the proposed
scheme is performed.
The rest of this paper is organized as follows. Section 2 is the proposed sys-
tem model. The proposed and existing schemes used to forecast electricity load
are discussed in Sect. 3. A detailed discussion on the simulation results of each
scheme is presented in Sect. 4. Section 5 concludes the paper along with the
future directions of our work.
2 System Model
Figure 3 presents a complete picture of our proposed system model. Forecasting
electricity load belongs to the paradigm of Time Series (TS) forecasting. Accu-
rate load forecasting using TS data is not an easy task. Therefore, in this paper,
we have first converted the TS electricity load data into a supervised learning
problem. Afterwards, training and testing is performed to forecast the electricity
load using the proposed model.
To convert the TS electricity load data into a supervised learning problem, we
have first extracted all days of the week separately from the historical load data.
All Mondays are extracted to form a Monday group. Similarly, all Tuesdays are
extracted to form a Tuesday group etc. This processes is repeated for all days.
1110 S. Khan et al.
After the formation of seven groups, we further divide each group into a train
and test dataset sub-group. Every train data sub-group contains load data of
eight days. Whereas, the test dataset sub-group contains the load data of the
ninth day. After the formation of train and test data sub-groups, these groups
are appended one after the other in such a way that all Mondays are followed by
all Tuesdays, Similarly all Tuesdays are followed by Wednesdays etc. Once all
the days are grouped into one set of train and test dataset, we create the input
and output variables such that output to Xi is Xi+1 and output of Xn is X1 .
The process of input and output variable formation is shown in Fig. 1.
Output
Sunday t1 t2 ... n-1 n t1
variables
In this paper, the dataset released by the Australian energy market operator is
used to forecast the electricity load of Victoria. The dataset is publicly available
at [1]. In this paper, the dataset used for forecasting contains load data from the
last week of April 2018 till the second week of July 2018. The load values are
recorded after every thirty minutes in a day and there are forty eight observations
in one day. Figure 2 shows the electricity load data used in this paper.
Electricity load (MW)
7500
5500
3500
0 500 1000 1500 2000 2500 3000
Time (30 min interval)
N
1
M AE = |Xi − Yi |2 (2)
N t=1
1 N
RM SE = |Xi − Yi |2 (3)
N t=1
Here Xi presents the actual load value, whereas, Yi presents the load forecast
values. To evaluate the forecasting results of a model, the lower the performance
metric parameter values, the better the results would be.
In this section we discussed the proposed and existing benchmark schemes used
to forecast the electricity load.
3.1 RNN
Among all the NN models, LSTM [21] is the most common type of RNN due
to its ability of memorizing information for a longer period of time. In LSTM,
every node act as a memory cell to store information. Furthermore, each node
has its own activation function. In order to process the input sequence, RNN
use the internal states. LSTM also have internal states, however, the difference
between other NN and LSTM is that in LSTM the old state becomes the input
to the next state along with the current input.
1112 S. Khan et al.
Start
Pre processing
Performance metrics
MAPE
Yes/No
MAE
RMSE
End
3.2 CNN
The most successful type of neural network to solve various problems in medical
imaging and computer vision is CNN [25]. It is due to the fact that CNN can
handle large amount of input sequences. Furthermore, it has the ability to extract
the hidden features in the input data. The architecture of CNN comprises of
convolution, fully connected, dense, flatten and max pooling layers. Commonly,
the deeper the CNN model, the higher the accuracy will be. In this paper, CNN
has one convolution, one dropout and two dense layers.
Electricity Load Forecasting for Each Day of Week Using Deep CNN 1113
3.3 ARIMA
It is one of the most commonly used statistical technique for TS forecasting.
ARIMA as a TS forecasting model is proposed by Box and Jenkins [22] as a
combination of Auto Regression (AR) and Moving Average (MA). The key idea
to combine AR and MA is to approximate the real world representation in a
more accurate and precise way. The main advantage of using ARIMA to forecast
electricity load is that it can accurately forecast the future load values if the TS
is a stationary series. However, if the TS is not stationary, then the input series
is converted to a stationary TS by taking the first or second difference.
3.4 ELM
Huang et al. in [23] and [24] proposed ELM, which a type of feed forward neu-
ral network. Due to its generalization and learning abilities, ELM as forecast-
ing model has gained popularity in several fields. The core advantage of ELM
includes: (i) parameters of the hidden layers do not require fine tunning: (ii)
biasness and the input weights to ELM are randomly assigned.
3.5 DCNN
We enhance the forecasting performance of CNN by adding extra layers. Increas-
ing the number of layers in CNN scale up the size and complexity of the proposed
DCNN. However, in this work, our focus is to minimize the difference between
the actual and forecast values. The proposed DCNN comprises of convolution
and dense layers. A dropout layer at 20% along with ReLU activation function.
10
MAPE Scores
8
0
ELM RNN DCNN CNN ARIMA
Figure 7 shows the electricity load forecast of Victoria for all days using CNN.
As shown in this figure, the performance of CNN is slightly better as compare to
RNN. However, the difference in the forecasted and actual electricity is still very
large. The bar plots in Figs. 4, 5 and 6 presents the performance errors metrics
of CNN. According to Table 1, CNN has the lowest error values in forecasting
the load of Thursday and Saturday. The MAPE, MAE and RMSE values for
CNN are 7.20%, 385.869 and 359.788 respectively. Whereas, in forecasting the
electricity load of Monday and Friday CNN has the worst performance metric
values.
500
MAE Scores
400
300
200
100
0
ELM RNN DCNN CNN ARIMA
500
RMSE Scores
400
300
200
100
0
ELM RNN DCNN CNN ARIMA
7500
Monday
5500
3500
7500
Tuesday
5500
3500
7500
Wednesday
5500
Elect ricit y load (MW)
3500
7500
Thursday
5500
3500
7500
Friday
5500
3500
7500
Sat urday
5500
3500
7500
Sunday
5500
3500
0 4 8 12 16 20 24
MAPE, MAE and RMSE in forecasting electricity load for Victoria are 2.15%,
138.771 and 116.417 respectively. Moreover, the lowest error values in forecasting
load using DCNN are achieved for Thursday. Similarly, the highest MAPE, MAE
and RMSE are 2.927%, 218.885 and 172.620. To sum up, by comparing the
usefulness of our proposed DCNN with the benchmark schemes, it can be said
that the proposed scheme has the best result in forecasting the electricity load
for all days.
to handle non linear and non stationary structure of the electricity load data.
In this paper, a DCNN is proposed to forecast the electricity load demand for
each day of week. Our experimental results illustrate that the proposed scheme
has the lowest MAPE, MAE and RMSE scores as compare to the benchmark
schemes. The proposed DCNN exceeds the existing benchmark schemes by scor-
ing the lowest MAPE, MAE and RMSE of 2.152%, 138.771 116.417. ELM has
the second best performance metric scores of 2.17, 152.037 and 117.390. ARIMA
achieved average results whereas, CNN and RNN has the worst performance
metric values. In future, we plan to investigate the impact of feature engineering
in forecasting the electricity load and price. Furthermore, to validate the effec-
tiveness of our proposed scheme, we plan to forecast the electricity load of other
states of Australia such as New South Wales and Queensland etc.
References
1. https://www.aemo.com.au/. Accessed 01 Jan 2019
2. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics
and energy consumption forecasting. Energies 11(2), 452 (2018)
3. Ahmad, T., Chen, H.: Short and medium-term forecasting of cooling and heating
load demand in building environment with data-mining based approaches. Energy
Build. 166, 460–476 (2018)
4. Chen, K., Chen, K., Wang, Q., He, Z., Hu, J., He, J.: Short-term load forecasting
with deep residual networks. IEEE Trans. Smart Grid (2018)
5. Jesus, L., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: Deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
6. Ryu, S., Noh, J., Kim, H.: Deep neural network based demand side short term load
forecasting. Energies 10(1), 3 (2016)
7. Liu, J., Li, C.: The short-term power load forecasting based on sperm whale algo-
rithm and wavelet least square support vector machine with DWT-IR for feature
selection. Sustainability 9(7), 1188 (2017)
8. Ashfaq, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and fast
converging short-term load forecasting model for industrial applications in a smart
grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
9. Huang, G.-B., Zhou, H., Ding, X., Zhang, R.: Extreme learning machine for regres-
sion and multiclass classification. IEEE Trans. Syst. Man Cybern. Part B (Cybern.)
42(2), 513–529 (2012)
10. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-
ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)
11. Zahid, M., Ahmed, F., Javaid, N., Abbasi, R.A., Kazmi, H.S.Z., Javaid, A., Bilal,
M., Akbar, M., Ilahi, M.: Electricity price and load forecasting using enhanced
convolutional neural network and enhanced support vector regression in smart
grids. Electronics 8(2), 122 (2019)
12. Fallah, S., Deo, R., Shojafar, M., Conti, M., Shamshirband, S.: Computational
intelligence approaches for energy load forecasting in smart energy management
grids: state of the art, future challenges, and research directions. Energies 11(3),
596 (2018)
13. Luo, J., Vong, C.-M., Wong, P.-K.: Sparse Bayesian extreme learning machine for
multi-classification. IEEE Trans. Neural Netw. Learn. Syst. 25(4), 836–843 (2014)
Electricity Load Forecasting for Each Day of Week Using Deep CNN 1119
14. Yu, J., Wang, S., Xi, L.: Evolving artificial neural networks using an improved PSO
and DPSO. Neurocomputing 71(4–6), 1054–1060 (2008)
15. Zeng, Y.-R., Zeng, Y., Choi, B., Wang, L.: Multifactor-influenced energy consump-
tion forecasting using enhanced back-propagation neural network. Energy 127,
381–396 (2017)
16. Saremi, S., Mirjalili, S.Z., Mirjalili, S.M.: Evolutionary population dynamics and
grey wolf optimizer. Neural Comput. Appl. 26(5), 1257–1263 (2015)
17. Ertugrul, Ö.F.: Forecasting electricity load by a novel recurrent extreme learning
machines approach. Int. J. Electr. Power Energy Syst. 78, 429–435 (2016)
18. Khan, M., Javaid, N., Naseem, A., Ahmed, S., Riaz, M., Akbar, M., Ilahi, M.:
Game theoretical demand response management and short-term load forecasting
by knowledge based systems on the basis of priority index. Electronics 7(12), 431
(2018)
19. Fan, C., Xiao, F., Zhao, Y.: A short-term building cooling load prediction method
using deep learning algorithms. Appl. Energy 195, 222–233 (2017)
20. Shi, H., Minghao, X., Li, R.: Deep learning for household load forecasting-a novel
pooling deep RNN. IEEE Trans. Smart Grid 9(5), 5271–5280 (2018)
21. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
22. Box, G.E.P., Jenkins, G.M., Reinsel, G.C., Ljung, G.M.: Time Series Analysis:
Forecasting and Control. Wiley, Hoboken (2015)
23. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning
scheme of feedforward neural networks. In: 2004 Proceedings of the IEEE Interna-
tional Joint Conference on Neural Networks, vol. 2, pp. 985–990. IEEE (2004)
24. Huang, G.-B., Chen, L., Siew, C.K.: Universal approximation using incremental
constructive feedforward networks with random hidden nodes. IEEE Trans. Neural
Netw. 17(4), 879–892 (2006)
25. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con-
volutional neural networks. In: Advances in Neural Information Processing Sys-
tems, pp. 1097–1105 (2012)
Short Term Load Forecasting
Using XGBoost
Abstract. For efficient use of smart grid, exact prediction about the in-
future coming load is of great importance to the utility. In this proposed
scheme initially we converted daily Australian energy market operator
load data to weekly data time series. Furthermore, we used eXtreme Gra-
dient Boosting (XGBoost) for extracting features from the data. After
feature selection we used XGBoost for the purpose of forecasting the elec-
tricity load for single time lag. XGBoost perform extremely well for time
series prediction with efficient computing time and memmory resources
usage. Our proposed scheme outperformed other schemes for mean aver-
age percentage error metric.
1 Introduction
Energy production and consumption difference minimization is a challenging
task these days. Efficient consumption of electricity is one good solution to this
problem. Researchers has done a lot of work in for introducing cost effective
and efficient energy utilization systems. Next generation Smart Grid (SG) is
the most attractive solution so far. SG is the integration of information and
communication technology in traditional grid which makes it intelligent power
grid supporting real-time information exchange between producer and consumer.
SG enables the energy efficiency optimization. More precisely, the SG needs an
accurate forecasting of the energy load for more productive application.
AN increasing attempt of deregulating strength markets to shape a more
dependable, green, and price-effective system with the aid of improving com-
petitions has been witnessed in world’s important economies [1,2]. In the lib-
eralized markets, the strength is commoditized and consequently its price is
dynamic. Due to the fee variant, pricing power correctly will become important
to generate profits, schedule strength productions, and plan load responses [3–
7]. The accurate power charge forecasting is helpful to decide the strength rate
and accordingly is precious. As liberalized electricity markets include types, day-
ahead and real-time [8], it’s miles meaningful to discuss both of the day-ahead
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1120–1131, 2019.
https://doi.org/10.1007/978-3-030-15035-8_108
STLF Using XGBoost 1121
and online forecasting of the power fee. The energy rate forecasting has been
vigorously studied inside the literature. From the application aspect, the fore-
casting of the energy rate in unique deregulated markets of essential economies
around the world has been stated [9–15].
The Rest of the paper is structured as defined next. The mostly used practi-
cle techniques for load forecasting are discussed in Sect. 2. The Sect. 3 enlightens
the different problems associated with the load prediction. The proposed model
for load prediction and the evaluation metrices are explained in Sect. 4. Exper-
imental results are depicted and highlighted in Sect. 5. Conclusion about the
work done in this paper is expressed in Sect. 6 at the end.
2 Related Work
Bayesian network forecasting, constant analysis of data through data mining and
unsupervised data accumulating for electricity consumption prediction. In [26]
authors proposed a short-lived electricity load prediction scheme for academic
buildings. This work used 2-stage forecasting analysis for the productive working
of their energy system. Energy consumption data is collected from different uni-
versities and moving average method is used for finding the energy load pattern
according to week day. Random Forest (RF) technique is used for forecasting
the daily energy load. RF performance is assessed using cross-validation on time
series.
González et al. predicted electricity price adopting functional time series
using a New Hilbertian ARMAX model in [27]. Suggested scheme has a lin-
ear regression structure, where functional variables are operated by functional
parameters. Where functional parameters are fundamental entities with linearly
combined kernels as sigmoid operations. Quasi-Newton model is used for param-
eters optimization in sigmoid which minimizes the sum of squared error. Data
integrity attacks affect the results of load prediction models i.e., artificial neural
network, multiple linear regression, support vector regression and fuzzy interac-
tion regression). Authors in [28] worked on exposing the consequences of these
attacks. We begin by simulating some knowledge integrity attacks through the
random injection of some multipliers that follow a traditional or uniform distri-
bution into the load series. Then, the four same load prognostication models are
used to generate one-year-ahead ex post purpose forecasts so as to supply a com-
parison of their forecast errors. The results show that the support vector regres-
sion model is most robust, followed closely by the multiple rectilinear regression
model, whereas the fuzzy interaction regression model is that the least sturdy of
the four. withal, all four models fail to supply satisfying forecasts once the size
of the info integrity attacks becomes giant. This presents a serious challenge to
each load forecasters and therefore the broader prognostication community: the
generation of correct forecasts beneath knowledge integrity attacks.
Dong et al. worked on the energy management in a microgrid. Bayesian-
optimization-algorithm (BOA) is used for a single SG using house. Authors
in [29] articulates the enhancement beyond the closed form equitable function
equation, and work out on it using BOA based data-driven technique. We can
consider the suggested technique as a black box function improving technique
as a whole. Furthermore, it has the ability to handle the microgrid working and
argument forecasting ambiguity.
We used various standards for the evaluation of our proposed prediction model
efficiency. The two most commonly used metrices for the measurement of predic-
tion accuracy are Mean Absolute Percentage Error (MAPE) and Mean Absolute
Error (MAE).
4.1.1 MAPE
The MAPE may be a live of prediction accuracy of a forecasting methodology for
constructing fitted statistic values in statistics, specifically in trend estimation.
it always expresses accuracy as a proportion of the error. as a result of this range
may be a percentage, it may be easier to know than the opposite statistics. The
MAPE is outlined as shown in (1). Here, and area unit the actual worth and
therefore the forecast worth, severally. Also, is the number of times discovered.
n
100 At − Ft
M AP E = (1)
n t=1 At
4.1.2 MAE
In statistics, the MAE is employed to measure however shut forecasts or predic-
tions area unit to the particular outcomes. It’s calculated by making a mean of
absolutely the variations between the prediction values and therefore the actual
ascertained values. The MAE is defined as shown in (2). Wherever nine is that
the prediction price, is the actual price.
n n
1 1
M AE = |fi − yi | = |ei | (2)
n i=1 n i=1
1126 R. A. Abbasi et al.
We used XGBoost for forecasting the load for a specific time lag in a week using
weekly data. Figure 5 shows the real load in the dataset and the XGBoost fore-
casted load. Here x-axis is depicting the time lags where y-axis is the load at
1128 R. A. Abbasi et al.
that specific time lag. Actual load is represented by the blue graph, where fore
casted load is represented by the green graph. We can see that the XGBoost
load prediction follows the real load at most of the time, however at some high
load instances XGBoost is not exactly following the real load. We can see that
XGBoost is not predicting well at the high loads.
The XGBoost load forecasting results for a time lag are displayed in Fig. 6.
We can see that XGBoost forecasting technique results in a low Mean Average
Percentage Error (MAPE), high accuracy and high Mean Average Error (MAE).
XGBoost load prediction resulted in a 10.08% MAPE, 97.21% accuracy and
88.90% MAE.
STLF Using XGBoost 1129
6 Conclusion
In this paper, we proposed a new scheme for electricity load forecasting. We con-
verted daily electricity load information into weekly load information. It increases
number of features available for predicting load for a lag variable. Then, we used
XGBoost, a recently dominant machine learning technique for time series pre-
diction, for feature selection from converted data. Once features are extracted
we train the model using XGBoost. After training we use trained model for load
prediction.
References
1. Mathaba, T., Xia, X., Zhang, J.: Analysing the economic benefit of electricity price
forecast in industrial load scheduling. Electr. Power Syst. Res. 116, 158–165 (2014)
2. Sarada, K., Bapiraju, V.: Comparison of day-ahead price forecasting in energy
market using Neural Network and Genetic Algorithm. In: Proceeding of the Inter-
national Conference on Smart Electric Grid, pp. 1–5 (2014)
3. Shafie-Khah, M., Moghaddam, M.P., Sheikh-El-Eslami, M.: Price forecasting of
day-ahead electricity markets using a hybrid forecast method. Energy Convers.
Manag. 52(5), 2165–2169 (2011)
4. Garcia, R.C., Contreras, J., Van Akkeren, M., Garcia, J.B.C.: A GARCH forecast-
ing model to predict day-ahead electricity prices. IEEE Trans. Power Syst. 20(2),
867–874 (2005)
5. Shahidehpour, M., Yamin, H., Li, Z.: Market overview in electric power systems. In:
Market Operations in Electric Power Systems, pp. 1–20. Wiley, New York (2002)
6. Ci-wei, G., Bompard, E., Napoli, R., Cheng, H.: Price forecast in the competitive
electricity market by support vector machine. Phys. A: Stat. Mech. Appl. 382(1),
98–113 (2007)
1130 R. A. Abbasi et al.
7. Cai, Y., Lin, J., Wan, C., Song, Y.: A stochastic Bi-level trading model for an
active distribution company with distributed generation and interruptible loads.
IET Renew. Power Gener. 11(2), 278–288 (2017)
8. Weron, R.: Electricity price forecasting: a review of the state-of-the-art with a look
into the future. Int. J. Forecast. 30(4), 1030–1081 (2014)
9. Hu, L., Taylor, G.: A novel hybrid technique for short-term electricity price fore-
casting in UK electricity markets. J. Int. Counc. Electr. Eng. 4(2), 114–120 (2014)
10. Voronin, S., Partanen, J.: Forecasting electricity price and demand using a hybrid
approach based on wavelet transform, ARIMA and neural networks. Int. J. Energy
Res. 38(5), 626–637 (2014)
11. Kou, P., Liang, D., Gao, L., Lou, J.: Probabilistic electricity price forecasting with
variational heteroscedastic Gaussian process and active learning. Energy Convers.
Manag. 89, 298–308 (2015)
12. Shrivastava, N.A., Panigrahi, B.K.: A hybrid wavelet-ELM based short term price
forecasting for electricity markets. Int. J. Electr. Power Energy Syst. 55, 41–50
(2014)
13. He, K., Xu, Y., Zou, Y., Tang, L.: Electricity price forecasts using a curvelet
denoising based approach. Phys. A: Stat. Mech. Appl. 425, 1–9 (2015)
14. Wan, C., Xu, Z., Wang, Y., Dong, Z.Y., Wong, K.P.: A hybrid approach for prob-
abilistic forecasting of electricity price. IEEE Trans. Smart Grid 5(1), 463–470
(2014)
15. Wan, C., Niu, M., Song, Y., Xu, Z.: Pareto optimal prediction intervals of electricity
price. IEEE Trans. Power Syst. 32(1), 817–819 (2017)
16. Liu, B., Nowotarski, J., Hong, T., Weron, R.: Probabilistic load forecasting via
quantile regression averaging on sister forecasts. IEEE Trans. Smart Grid 8(2), 1
(2017). https://doi.org/10.1109/tsg.2015.2437877
17. Wanga, P., Liu, B., Hongb, T.: Electric load forecasting with recency effect: a big
data approach. Int. J. Forecast. 32, 585–597 (2016)
18. Chou, J.-S., Ngo, N.-T.: Smart grid data analytics framework for increasing energy
savings in residential buildings. Autom. Constr. 72, 247–257 (2016)
19. Ludwig, N., Feuerriegel, S., Neumann, D.: Putting big data analytics to work:
feature selection for forecasting electricity prices using the LASSO and random
forests. ISSN 1246-0125 (Print) 2116-7052
20. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
21. Wang, L., Zhang, Z., Chen, J.: Short-term electricity price forecasting with stacked
denoising autoencoders. IEEE Trans. Power Syst. 32(4), 2673–2681 (2017)
22. Lagoa, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
tricity prices in Europe: the importance of considering market integration. Appl.
Energy 211(1), 890–903 (2018)
23. Raviv, E., Bouwman, K.E., van Dijk, D.: Forecasting day-ahead electricity prices:
utilizing hourly prices. Energy Econ. 50, 227–239 (2015)
24. Xiao, L., Jianzhou Wang, R., Hou, J.W.: A combined model based on data pre-
analysis and weight coefficients optimization for electrical load forecasting. Energy
82(15), 524–549 (2015)
25. Singh, S., Yassine, A.: Big data mining of energy time series for behavioral analytics
and energy consumption forecasting. Energies 11, 452 (2018)
26. Moon, J., Kim, K.-H., Kim, Y., Hwang, E.: A short-term electric load forecasting
scheme using 2-stage predictive analytics. In: 2018 IEEE International Conference
on Big Data and Smart Computing (2018)
STLF Using XGBoost 1131
27. González, J.P., San Roque, A.M., Perez, E.A.: Forecasting functional time series
with a new Hilbertian ARMAX model: application to electricity price forecasting.
IEEE Trans. Power Syst. 33(1), 545–556 (2018)
28. Luo, J., Hong, T., Fang, S.-C.: Benchmarking robustness of load forecasting models
under data integrity attacks. Int. J. Forecast. 34, 89–104 (2018)
29. Dong, G., Chen, Z.: Data driven energy management in a home microgrid based
on Bayesian optimal algorithm. IEEE Trans. Ind. Inform
30. Fausett, L.: Fundamentals of Neural Networks: Architectures, Algorithms, and
Applications. Pearson Education, Delhi (2006)
31. Shekhar, S., Amin, M.B.: Generalization by neural networks. IEEE Trans. Knowl.
Data Eng. 4, 177–185 (1992)
Electricity Price Forecasting in Smart
Grid: A Novel E-CNN Model
Abstract. The vital part of the smart grid is electricity price forecast-
ing because it makes grid cost saving. Although, existing systems for
price forecasting may be challenging to manage with enormous price
data in the grid. As repetition from the feature cannot be avoided and
an integrated system is needed for regulating the plans in price. To han-
dle this problem, a new price forecasting system is developed. This pro-
posed model particularly integrated with three systems. Initially, fea-
tures are selected from the random data by combining the Mutual Infor-
mation (MI) and Random Forest (RF). The Grey Correlation Analysis
(GCA) is used to remove the redundancy from the selected features.
Secondly, the Recursive Feature Elimination (RFE) scheme is used to
reduce the dimensions. Finally, classification is done based on Enhanced-
Convolutional Neural Network (E-CNN) classifier to forecast the price.
The simulation results show that our accuracy of the proposed system is
higher than existing benchmark schemes.
1 Background
The main objective of the smart grid is to reduce the energy peak load and
to balance the space between power supply and demand. Electricity consumers
can interact in the working of smart grid to conserve the energy and shift the
load. As for users, price forecasting is important because knowing the price they
can shift the load On and Off. In the last couple of years, high power demand
in domestic areas has been observed due to an expansion in electrical devices.
The power demand has extended up to 28% due to the rise in energy consump-
tion. Many researchers working on the residential area to control and manage
the household appliances. In [1], authors proposed a novel Memory Updation
Heuristic Scheme (MUHS), which efficiently schedule the appliances from on-
peak to off-peak time slots. We studied the literature work in price forecasting
and feature selection. Electricity price forecasting is done by time series pre-
diction and machine learning algorithms. In [2], authors proposed a model to
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1132–1144, 2019.
https://doi.org/10.1007/978-3-030-15035-8_109
Electricity Price Forecasting 1133
foretell the day ahead power price according to load demand and temperature,
with the usage of neural network design and evaluation of the unique spectrum.
The authors in [3], presented a probabilistic system to determine hourly price
rate, where the bootstrapping framework is used for examining uncertainty and
induced extreme learning machine strategy is introduced for wavelet neural net-
works. In [4], authors proposed a periodically Autoregressive Moving Average
Hilbertian (AMAH) design to determine the changeable average values in prac-
tical time series, which can be implemented to price prediction.
In [5], authors proposed a deep neural system model in the regional market to
increase the prediction accuracy, Furthermore, a new feature selection algorithm
is proposed to measure the feature importance by utilizing Bayesian optimiza-
tion. To improve accuracy, researchers forecast prices from two local markets
simultaneously. In paper [6], a novel price forecasting framework is proposed,
which consist of four deep learning models for price forecasting to increase pre-
diction accuracy. To examine the model extensively, the authors evaluated 27
algorithms and stated that our proposed deep model beats the benchmarks.
Prior researches mainly concentrate on feature algorithms or classifiers scheme,
where conventional classifiers, e.g., Decision Tree (DT) are used in [7].
Yet, DT normally faces the over fit obstacle, which indicates the DT shows
well in training but not in prediction. Luckily, with the advancement of artificial
intelligence, several artificial intelligence predictions algorithms, including Elman
Neural Network (ENN), Generalized Regression Neural Network (GRNN), Arti-
ficial Neural Networks (ANNs), Wavelet Neural Network (WNN) [8], etc., Some
classifiers like Extreme Learning Machine (ELM) [9] and Least Squares Support
Vector Machine (LSSVM) [10] has been improved in forecasting domains, such
as air conditions advanced warning [11]. However, the literature contains many
techniques for determining electricity price. Every technique has their own ben-
efits and limitations, hence cannot always accomplish the desired achievement.
Clearly, the analytical systems have a bad extrapolation result and tight fore-
casting range, with a considerable dependency on data. Furthermore, these are
proper for linear data and inappropriate for data featuring variations and noise.
The Grey Design (GD) is proper for data by having an exponential curve but
improper to follow the trend of a waved series [12]. From a common aspect,
numerous applications might start modifying 1-D Convolutional Neural Net-
work (CNN) for linear data used for time series forecasting [13]. Short term load
and price forecasting is also propose in [14,15]. A novel game theory approach
was also proposed for load forecasting in [16]. Short term load forecasting in
industrial area is also proposed in [17]. In this research work, an Enhanced-CNN
is proposed for price forecasting. We have applied tenser flow to develop our
system model.
smart grid to improve its efficiency. The smart grid has a bidirectional digital
communication channel between end users and utility. The aim of the smart
grid is to reduce the cost and burden on utility. This can be done through price
forecasting. Several deep layer models are proposed in the field of time series
and energy-related sides to bring the best system model in the area of elec-
tricity price forecasting. Specifically, many neural networks models have been
presented which have limitations of hidden layer and accuracy e.g., Recurrent
Neural Network (RNN) [18] and Multilayer Perceptron (MLP) [19].
3 Proposed Solution
Based on above limitations, the main objective and focus of this paper are to
develop a deep layer model that can efficiently and accurately predict the next
day prices. In this paper a deep learning system model is developed which is
described as follows:
We presented an integrated price forecasting structure to obtain accurate big
data prediction in the smart grid. The primary effort in this paper is categorized
in hybrid feature selection, feature extraction, and classification based on E-CNN
are deployed in our system model.
To build this system model, a new Grey Correlation Analysis (GCA) feature
selection with Mutual Information (MI) and Random Forest (RF) are used to
measure the feature importance. In order to reduce the redundancy from the
selected features, we used Recursive Feature Elimination (RFE) algorithm. Fur-
thermore, we also design E-CNN to increase the accuracy than existing bench-
marks.
The efficiency of our system is evaluated through many simulations with ISO
New England real price data 2016. The results reveal that our system model
outperformed than benchmark schemes.
4 System Model
Our system model consists of three stages i.e., GCA based MI and RF feature
selection, RFE based feature extraction and E-CNN based classification.
Decision Tree takes the data and creates a sequence of conditions into a tree
formation. The nodes and roots of decision trees are comprised of attributes that
separate the data into various features according to conditions. After that classes
are named with node labels true or false. When the decision tree is developed,
the classification of a test set becomes easy. Beginning the node, we select the
relevant test set to record and follow the outcome. In order to foretell the class
label and to trace its way in a decision tree for the record, the route ended at a
leaf with no labeled.
SVM classifier is used to reduce the risk function. SVM super parameters and
risk function have a good relationship. Some SVM parameters are insensitive loss
function, cost penalty, and kernel. However, the tuning of these SVM parameters
is important and difficult task to achieve high accuracy and efficiency. The sim-
ple way to tune SVM parameters is with cross-validation and gradient descent.
Despite of these two tuning methods, computational complexity is increased
which reduce the performance of SVM.
CNN is a sample of network topology, generally practiced to use for deep learn-
ing systems. CNN is a controlled neural network. Similar to the conventional
structure of a neural network which consists of input, hidden and finally output
layers, the CNN also comprises of these features and input of one layer is the
output of another layer. Still CNN has some good features like pooling layer
and full connected layer. Of course, The CNN have more hidden layers than
conventional neural network.
Electricity Price Forecasting 1137
6 Proposed Scheme
In this paper, we introduce a new approach by linking a deep learning design with
conventional machine learning systems to manage short term price forecasting.
Particularly, we utilize the E-CNN with clustering models and enhanced it to
obtain more accuracy in prediction. The conv1d function is used to feed the Con-
volutional layer with 1D data. As our model is linear so we used Relu activation
function to bypass the gradient vanishing problem.
0(x ≤ 0)
Relu(x) = (1)
x(x > 0)
Where x represents the convolution computation in layer and we selected the
0.1 value for tuning in our scenario. The process has two stages comprising of
four steps. The first step, we preprocess the input raw data of price to remove
redundancy from the features. In second step, we divide the whole raw data into
training and testing sets. In the third stage, an E-CNN is prepared with all of
these sets to develop the prediction model. The last stage validation is conducted
to evaluate the prediction model accuracy. The E-CNN further comprises of
two stages. In the first stage, the training data is preprocessed with traditional
feature selection and extraction algorithms and then E-CNN based system model
is constructed. In the second stage, testing data is taken as an input and ensemble
with E-CNN based system model to achieve better forecasting result. The four
layer E-CNN which comprise of Convolutional layer, activation function, max
pooling and fully connected layer showing in Fig. 2. Furthermore, by tuning all
these four layers we achieved the desired accuracy for price forecasting.
In E-CNN (Wi = 1, 2, ..L), represents the layer weights and n represents learning
rate. In this algorithm, randomly one weight is selected and reduce its learning
possibility to zero. When the training is over using back propagation and epocs,
1138 W. Ahmad et al.
the value of Wi is then modified to original n value. The details are described in
Algorithm 1. For the counter loop, the sub epocs is used which is set to zero in
nk. The predefined set values are Ci(i = 1, 2.., L). The benefit of using training
data is to train the neural network efficiency.
Algorithm for E-CNN
7 Simulation Setup
In order to examine the efficiency of our proposed model, we developed a python
framework according to the proposed system. Throughout the experiment, the
simulator is operating on HP system with Core i5, 6 GB RAM and 500 GB hard
disk. ISO New England Control Area (ISO NE-CA) 2016 [20] electricity price
information is used as the input for our proposed framework. The real-world
new England electricity price data consists of over 1000 records. The results of
the simulation are described as follows:
• 1. GCA based Hybrid feature selection,
• 2. RFE performance score,
• 3. E-CNN performance compared with benchmark schemes,
• 4. E-CNN robustness compared with benchmark schemes.
Electricity Price Forecasting 1139
The aim of RFE is to choose the features by recursively analyzing the set of fea-
tures. In the beginning, the estimator is practiced on the original list of features
and the quality of an individual feature is acquired. Then, the small primary
features are extracted from the current list of features. This method is recur-
sively iterated on the clipped list until the required list of features to choose is
finally accomplished. To obtain the optimal figure of features, cross-validation
is employed with RFE to score various feature. The Fig. 5 presents a good RFE
cross-validated curve at 0.876 points and five important features are obtained.
This curve slowly declines in precision as the less important features are contin-
ued into the system model.
We evaluate the E-CNN with four benchmark classifiers i.e., RF, DT, SVM
and CNN. The Accuracy bars are shown in Fig. 6, which indicate that E-CNN
beats the benchmark schemes and achieved the accuracy of 97.03%. The E-
CNN curve line in Fig. 7 beats the actual and benchmark schemes by forecasting
less electricity price. The E-CNN shows satisfying results by determining less
electricity price. The performance among the benchmark algorithms indicates
that each module simulated in our framework can efficiently boost the accuracy.
The E-CNN efficiently decrease the inconsistency and repetition between the
features in order to increase the price forecasting accuracy.
Electricity Price Forecasting 1141
data which is not part of the same feature set. We selected the first method
and the solid lines in Fig. 8 showing the precision of algorithms. Basically, the
noise is produced by the failure of sensors. As the results showing that our
proposed E-CNN is more reliable as compared to other three existing algorithms.
In our proposed framework, each feature has a short impact of noise towards the
accuracy, because less important features are discarded initially in selection and
extraction phase from the feature set.
9 Conclusion
References
1. Ahmad, W., Javaid, N., Sajjad, M.Z., Awan, T., Amir, M.: A new memory upda-
tion heuristic scheme for energy management in smart grid (2018)
2. Varshney, H., Sharma, A., Kumar, R.: A hybrid approach to price forecasting incor-
porating exogenous variables for a day ahead electricity market. In: IEEE Interna-
tional Conference on Power Electronics, Intelligent Control and Energy Systems
(ICPEICES), pp. 1–6. IEEE (2017)
3. Rafiei, M., Niknam, T., Khooban, M.-H.: Probabilistic forecasting of hourly elec-
tricity price by generalization of ELM for usage in improved wavelet neural net-
work. IEEE Trans. Ind. Inform. 13(1), 71–79 (2017)
4. González, J.P., San Roque, A.M., Perez, E.A.: Forecasting functional time series
with a new Hilbertian ARMAX model: application to electricity price forecasting.
IEEE Trans. Power Syst. 33(1), 545–556 (2018)
5. Lago, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
tricity prices in Europe: the importance of considering market integration. Appl.
Energy 211, 890–903 (2018)
6. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
7. Wang, K., Yu, J., Yu, Y., Qian, Y., Zeng, D., Guo, S., Xiang, Y., Wu, J.: A survey
on energy internet: architecture, approach and emerging technologies. IEEE Syst.
J. PP(99), 1 (2017)
1144 W. Ahmad et al.
8. Wang, J., Yang, W., Du, P., Niu, T.: A novel hybrid forecasting system of wind
speed based on a newly developed multi-objective sine cosine algorithm. Energy
Convers. Manag. 163, 134–150 (2018)
9. Wang, J., Niu, T., Lu, H., Guo, Z., Yang, W., Du, P.: An analysis-forecast system
for uncertainty modeling of wind speed: a case study of large-scale wind farms.
Appl. Energy 211, 492–512 (2018)
10. Li, C., Xiao, Z., Xia, X., Zou, W., Zhang, C.: A hybrid model based on synchronous
optimisation for multi-step short-term wind speed forecasting. Appl. Energy 215,
131–144 (2018)
11. Xu, Y., Yang, W., Wang, J.: Air quality early-warning system for cities in China.
Atmos. Environ. 148, 239–257 (2017)
12. Yang, W., Wang, J., Wang, R.: Research and application of a novel hybrid model
based on data selection and artificial intelligence algorithm for short term load
forecasting. Entropy 19(2), 52 (2017)
13. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-
ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)
14. Nayab, A., Javaid, N.: Load and price forecasting in smart grids using enhanced
support vector machine
15. Khan, M., Javaid, N., Naseem, A., Ahmed, S., Riaz, M., Akbar, M., Ilahi, M.:
Game theoretical demand response management and short-term load forecasting
by knowledge based systems on the basis of priority index. Electronics 7(12), 431
(2018)
16. Ahmad, A., Javaid, N., Guizani, M., Alrajeh, N., Khan, Z.A.: An accurate and
fast converging short-term load forecasting model for industrial applications in a
smart grid. IEEE Trans. Ind. Inform. 13(5), 2587–2596 (2017)
17. Wang, Z., Yan, W., Oates, T.: Time series classification from scratch with deep
neural networks: a strong baseline. In: 2017 International Joint Conference on
Neural Networks (IJCNN), pp. 1578–1585. IEEE (2017)
18. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge
(2016)
19. Panapakidis, I.P., Dagoumas, A.S.: Day-ahead electricity price forecasting via the
application of artificial neural network based models. Appl. Energy 172, 132–151
(2017)
20. ISO New England Energy Offer Data [Online] (2016). www.iso-ne.com/isoexpress/
web/reports/pricing/-/tree/dayahead-energy-offer-data
Electricity Price Prediction by Enhanced
Combination of Autoregression Moving
Average and Kernal Extreme
Learing Machine
1 Background
In the past decade of electricity market, the volatile electricity prices became a
complicated phenomenon along with few characteristics and important concep-
tion. Market’s managers have to ensure the stability of forecasting for the power
market. In power, money contribution decisions and transmission inflation, pre-
dictions of electricity price play an important role. However, it is a complicated
task to predict accurate price of electricity because of many segregated fea-
tures, like as high volatility, nonlinearity, multiple seasonality, etc. In past few of
c Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): WAINA 2019, AISC 927, pp. 1145–1156, 2019.
https://doi.org/10.1007/978-3-030-15035-8_110
1146 S. M. Shuja et al.
In this section of the paper, we describe in detail proposed system model for
method develop to forecast the electricity price and is shown in Fig. 1. It is very
difficult to forecast the accurate electricity price, because of many circumstances
including weather, load and previous predicted prices. In addition, balance is
required between demand and supply as electricity cannot be stored due to the
inelastic nature of electricity shorter comes and unstable nature of generation
[15], complexity is produced in the behavior of electricity price. In this paper, we
1148 S. M. Shuja et al.
In this section, we discuss the techniques used for the predictions of electricity
price.
1150 S. M. Shuja et al.
This model is used in many areas of researches with respect of simple ELM and
ELM comes out of SFLN. Its responsiveness behavior toward non-stationary and
volatile series is due to fast learning and forecasting ability in price. In ELM,
hidden layer’s parameters do not need tuning. In [17], Huang proposed the kernel
based ELM. Kernel matrix in ELM is used, when feature mapping is unknown
for the user. The user has not specified hidden nodes and the hidden layer feature
mapping. In the stable kernel function is changing the randomness of ELM that
why KELM is better in generalization.
This is the model based on random time series used for predictive analysis [18].
Prediction analysis of time series values consist of three types: AR, MA and
ARMA. The ARMA model implementation is based on following steps:
First, stationarity analysis is performed through the ADF test in the wavelet
decomposition series of NSW data. The ADF test results show that the ADF-
statistic values of detail series, D1, D2, D3 and D4 are smaller than 1% critical
level, thus their unit root hypothesis is rejected. However, ADF-statistic value
of an approximate series A greater than 10% critical level, thus its unit root
hypothesis is not rejected. Stationarity analysis indicates that A is identified as
non-stationary series and other detail series are stationary series.
Second, we applied ARMA-model on stationary series contain four detail
series and KELM-model of non-stationary series having one approximate series.
After that, we enhanced the results for both of the models along with the sum
of their predicted price values.
To elaborate the efficiency of our proposed method for prediction, data from
NSW is tested for January month of the year 2018 and the three type errors
MAPE, MSE and RMSE are calculated. We compare the results of our enhanced
proposed scheme with ARMA-model and KELM-model for errors. NSW electric-
ity price forecasted data is predicted by ARMA-model is shown in Fig. 2. The
peak of predicted values depicts the accuracy of single ARMA-model.
The prediction of enhanced combined prediction for NSW data of electricity
price is depicted in the Fig. 4. Peaks of graphs show that when ARMA and KELM
for were used separately and were not combined, then the accuracy value price
prediction was less than the proposed scheme. Complex price feature cannot be
resolved by single KELM and ARMA. ARMA is not capable to solve the non-
linear issues, because its forecasted values are not more accurate than KELM.
In Fig. 3, single KELM-model based electricity forecasting of price is shown.
Fig. 6. MSE error in percentage comparison of our enhanced combined prediction with
ARMA and KELM
Enhanced Combined Price Prediction 1155
5 Conclusion
Due to uncontrolled electricity market, predicting the price of electricity has
become essential to power users and supplier. Implementing an efficient and an
accurate prediction model has become a very important role in the electricity
scope. In this paper, a new enhanced method for prediction of electricity price
combining KELM and ARMA forecasted output is proposed. Electricity dereg-
ulated price series is transformed and decomposed into different series by WT,
the KELM-model forecast the non-stationary part of decomposed price, while
ARMA-model is used to predict the stationary part of the price. Final forecasting
of electricity price is the combination of KELM-model and ARMA-model. NSW
market data is used to verify the forecasting ability of our proposed enhanced
combine method. This paper is completed depends on the improvement of fore-
casting accuracy through enhanced method. Although some prices patterns of a
specific time period are not matched, the proposed method improved the match-
ing feature for prediction of electricity price. In future, load and weather should
be considered for the prediction accuracy through our proposed method.
References
1. Cerjan, M., Krzelj, I., Vidak, M., Delimar, M.: A literature review with statistical
analysis of electricity price forecasting methods. In: 2013 IEEE EUROCON, pp.
756–763. IEEE (2013)
2. Weron, R.: Electricity price forecasting: a review of the state-of-the-art with a look
into the future. Int. J. Forecast. 30, 1030–1081 (2014)
3. Chan, S.-C., Tsui, K.M., Wu, H., Hou, Y., Wu, Y.-C., Wu, F.F.: Load/price fore-
casting and managing demand response for smart grids: methodologies and chal-
lenges. IEEE Signal Process. Mag. 29, 68–85 (2012)
1156 S. M. Shuja et al.
4. Huang, G.-B., Zhu, Q.-Y., Siew, C.-K.: Extreme learning machine: a new learning
scheme of feedforward neural networks. In: 2004 Proceedings 2004 IEEE Interna-
tional Joint Conference on Neural Networks, pp. 985–990. IEEE (2004)
5. Lago, J., De Ridder, F., De Schutter, B.: Forecasting spot electricity prices: deep
learning approaches and empirical comparison of traditional algorithms. Appl.
Energy 221, 386–405 (2018)
6. Lago, J., De Ridder, F., Vrancx, P., De Schutter, B.: Forecasting day-ahead elec-
tricity prices in Europe: the importance of considering market integration. Appl.
Energy 211, 890–903 (2018)
7. Loi, T.S.A., Le Ng, J.: Anticipating electricity prices for future needs-Implications
for liberalised retail markets. Appl. Energy 212, 244–264 (2018)
8. Bento, P.M.R., Pombo, J.A.N., Calado, M.R.A., Mariano, S.J.P.S.: A bat opti-
mized neural network and wavelet transform approach for short-term price fore-
casting. Appl. Energy 210, 88–97 (2018)
9. Shuja, S.M., Javaid, N., Rafique, M.Z.: Towards Efficient Scheduling of Smart
Appliances for Energy Management by Candidate Solution Updation Algorithm
(CSUA) in Smart Grid
10. Chitsaz, H., Zamani-Dehkordi, P., Zareipour, H., Parikh, P.P.: Electricity price
forecasting for operational scheduling of behind-the-meter storage systems. IEEE
Trans. Smart Grid 9(6), 6612–6622 (2018)
11. Raa, P., Vilar, J., Aneiros, G.: On the use of functional additive models for elec-
tricity demand and price prediction. IEEE Access 6, 9603–9613 (2018)
12. Alanis, A.Y.: Electricity prices forecasting using artificial neural networks. IEEE
Lat. Am. Trans. 16(1), 105–111 (2018)
13. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.: Short-term load forecast-
ing in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019)
14. Khan, M., Javaid, N., Naseem, A., Ahmed, S., Riaz, M., Akbar, M., Ilahi, M.:
Game theoretical demand response management and short-term load forecasting
by knowledge based systems on the basis of priority index. Electronics 7(12), 431
(2018)
15. Esther, B.P., Kumar, K.S.: A survey on residential demand side management archi-
tecture, approaches, optimization models and methods. Renew. Sustain. Energy
Rev. 59, 342–351 (2016)
16. AEMO. http://www.aemo.com.au/Electricity/Data/Price-and-Demand/Aggrega
ted-Price-and-Demand-Data-Files
17. Huang, G.-B., Chen, L.: Enhanced random search based incremental extreme learn-
ing machine. Neurocomputing 71, 34608 (2008)
18. Box, G.E., Jenkins, G.M., Reinsel, G.C.: Time Series Analysis: Forecasting and
Control. Wiley, Hoboken (2011)
Prediction of Building Energy
Consumption Using Enhance
Convolutional Neural Network
1 Background
2 Problem Statement
In this paper, we focus the problem of electricity load forecasting. The aim of our
model is to predict the exact load forecast with maximum amount of data from
smart grid. To tackle load forecasting problem, four classifiers are applied to
predict the electricity load. CNN is a classifier that splits the data into training
and testing phase. CNN is efficient classifier, however, following problems are
need to answered for good accuracy of electricity load forecasting.
3 System Model
Our proposed system model consist of four parts: normalizing, training testing
and prediction. Four classifiers with tuned parameters as shown in Fig. 1.
1160 H. M. Faisal et al.
System load data is collected for years 2015 to 2017 from ISO-NECA. Data is
divided on monthly basis and each years similar month data is grouped together
e.g. January 2015, January 2016 and January 2017 data is grouped together.
First three weeks of any month are used for the training process and the last
week is for testing. Data is normalized with maximum impact values and divided
into three major parts: training, testing and validation. To investigate the per-
formance of the four classifiers, we used three performance evaluation errors:
mean square percentage error (MSPE), root mean square error (RMSE) and the
mean absolute percentage error (MAPE) which can be calculated respectively as
TM
1
M SP E = |(Av − Fv )| (1)
T tm=1
TM
1
RM SE =
2
(Av − Fv ) (2)
T tm=1
TM
1 Av
M AP E = 100 (3)
T tm=1 Fv
where Av is the observed test value at time tm and Fv is the forecasted value
at time tm.
4 Proposed Scheme
In this section, we will discuss RFE and enhance CNN classifiers.
Prediction of Building Energy Consumption 1161
4.1 RFE
RFE method is used for feature selection. Feature selection method fits a model
and discards the weakest feature until the specified number of features is reached.
Fig. 6. CNN
Fig. 9. NN
6 Conclusion
Proposed deep learning shown its great learning prediction qualities. The
proposed model consists of two key parts; feature selection and classifiers. A
combination of two different techniques are used for feature selection. The pro-
posed model shows the best load prediction results in building sector. Pro-
posed model provided a novel prediction model for building energy consumption.
Comparison results indicate that the enhance CNN method performances better
than the existing machine learning methods.
References
1. Streimikiene, S.: Residential energy consumption trends, main drivers and policies
in Lithuania. Renew. Sustain. Energy Rev. 35, 285–293 (2014)
2. Ugursal, V.I.: Energy consumption, associated questions and some answers. Appl.
Energy 130, 783–792 (2014)
3. Hua, C., Lee, W.L., Wang, X.: Energy assessment of office buildings in China using
China building energy codes and LEED 2.2. Energy Build. 86, 514–524 (2015)
Prediction of Building Energy Consumption 1167
4. Zuo, J., Zhao, Z.Y.: Green building research-current status and future agenda: a
review. Renew. Sustain. Energy Rev. 30, 271–281 (2014)
5. Daut, M.A.M., Hassan, M.Y., Abdullah, H., Rahman, H.A., Abdullah, M.P.,
Hussin, F.: Building electrical energy consumption forecasting analysis using con-
ventional and artificial intelligence methods: a review. Renew. Sustain. Energy
Rev. 70, 1108–1118 (2017)
6. Li, K., Hu, C., Liu, G., Xue, W.: Building’s electricity consumption prediction using
optimized artificial neural networks and principal component analysis. Energy
Build. 108, 106–113 (2015)
7. Pombeiro, H., Santos, R., Carreira, P., Silva, C., Sousa, J.M.C.: Comparative
assessment of low-complexity models to predict electricity consumption in an insti-
tutional building: linear regression vs. fuzzy modeling vs. neural networks. Energy
Build. 146, 141–151 (2017)
8. Jimenez, M.J., Heras, M.R.: Application of multi-output ARX models for estima-
tion of the U and g values of building components in outdoor testing. Sol. Energy
79, 302–310 (2005)
9. Kimbara, A., Kurosu, S., Endo, R., Kamimura, K., Matsuba, T., Yamada, A.: On-
line prediction for load profile of an air-conditioning system. Ashrae Trans. 101,
198–207 (1995)
10. Newsham, G.R., Birt, B.J.: Building-level occupancy data to improve ARIMA-
based electricity use forecasts. In: Proceedings of the 2nd ACM Workshop on
Embedded Sensing Systems for Energy-Efficiency in Building, Zurich, Switzerland,
pp. 13–18, 2 November 2010
11. Aydinalp-Koksal, M., Ugursal, V.I.: Comparison of neural network, conditional
demand analysis, and engineering approaches for modeling end-use energy con-
sumption in the residential sector. Appl. Energy 85, 271–296 (2008)
12. Hsu, D.: Comparison of integrated clustering methods for accurate and stable
prediction of building energy consumption data. Appl. Energy 160, 153–163 (2015)
13. Alvarez, F.M., Troncoso, A., Riquelme, J.C., Ruiz, J.S.A.: Energy time series fore-
casting based on pattern sequence similarity. IEEE Trans. Knowl. Data Eng. 23,
1230–1243 (2011)
14. Pérez-Chacón, R., Talavera-Llames, R.L., Martinez-Alvarez, F., Troncoso, A.:
Finding electric energy consumption patterns in big time series data. In: Proceed-
ings of the 13th International Conference Distributed Computing and Artificial
Intelligence, Sevilla, Spain, 1–3 June 2016, pp. 231–238. Springer, Cham (2016)
15. Martı́nez-Álvarez, F., Troncoso, A., Asencio-Cortés, G., Riquelme, J.C.: A survey
on data mining techniques applied to electricity-related time series forecasting.
Energies 8, 13162–13193 (2015)
16. Tso, G.K.F., Yau, K.K.W.: Predicting electricity energy consumption: a compari-
son of regression analysis, decision tree and neural networks. Energy 32, 1761–1768
(2007)
17. Ahmad, M.W., Mourshed, M., Rezgui, Y.: Trees vs. neurons: comparison between
random forest and ANN for high-resolution prediction of building energy consump-
tion. Energy Build. 147, 77–89 (2017)
18. Paudel, S., Elmitri, M., Couturier, S., Nguyen, P.H., Kamphuis, R., Lacarriàre, B.,
Corre, O.L.: A relevant data selection method for energy consumption prediction of
low energy building based on support vector machine. Energy Build. 138, 240–256
(2017)
19. Mena, R., Rodrı́guez, F., Castilla, M., Arahal, M.R.: A prediction model based
on neural networks for the energy consumption of a bioclimatic building. Energy
Build. 82, 142–155 (2014)
1168 H. M. Faisal et al.
20. Biswas, M.A.R., Robinson, M.D., Fumo, N.: Prediction of residential building
energy consumption: a neural network approach. Energy 117, 84–92 (2016)
21. Ahmad, A., Javaid, N., Mateen, A., Awais, M., Khan, Z.A.: Short-term load fore-
casting in smart grids: an intelligent modular approach. Energies 12(1), 164 (2019).
https://doi.org/10.3390/en12010164. ISSN: 1996-1073
22. Zahid, M., Ahmed, F., Javaid, N., Abbasi, R.A., Kazmi, S.Z., Javaid, A., Bilal,
M., Akbar, M., Ilahi, M.: Electricity price and load forecasting using enhanced
convolutional neural network and enhanced support vector regression in smart
grids. Electronics (2019). EISSN 2079-9292. (IF = 2.110, Q2)
Author Index
Khan, Sajjad, 67, 80, 385, 398, 1107, 1145 Mohebi, Nicolay, 1020
Khan, Wahab, 1084 Moscato, Francesco, 815, 822
Khan, Wazir Zada, 592 Moscato, Vincenzo, 880
Khan, Zahoor Ali, 67, 1120 Mostarda, Leonardo, 734
Khan, Zain Ahmad, 398, 1071 Mubarak, Hasnain, 371
Kirsal Ever, Yoney, 713 Mubin, Omar, 907
Ko, Chien-Chuan, 768 Murakami, Shunsuke, 161
Kolici, Vladi, 205 Murano, Aniello, 870
Kuang, Hanbao, 1007
Kulla, Elis, 351
N
Kurita, Takeru, 604
Nabeta, Shinichi, 501
Nagatomo, Makoto, 683
L
Naji, Mohamad, 923
Lassoued, Narjes, 244
Nakamura, Shigenari, 570, 612
Lee, Hui-Lan, 441
Nakasaki, Shogo, 215
Lee, JooYoung, 116
Nandi, Sukumar, 129, 651
Li, Hui, 757
Nandi, Sunit Kumar, 129, 651
Li, Kin Fun, 334
Navarra, Alfredo, 734
Li, Man, 286
Naz, Aqdas, 1097
Li, Peng, 35, 561
Nung, Lin-Hui, 451
Liang, Xiaobing, 633
Lin, Chi-Yi, 471
O
Lin, Chun-Li, 776
Oda, Tetsuya, 362
Lin, Hsuan-Yu, 776
Odagiri, Kazuya, 44
Lin, Iuon-Chang, 459
Ogura, Naohiro, 57
Lin, Jung-Shan, 451
Ohmori, Motoyuki, 151
Lin, Pei-Shan, 793
Ohyanagi, Tatsuya, 516
Lin, Pei-Yu, 471
Oikawa, Takanori, 673
Lin, Wei-Tsen, 451
Okazaki, Naonobu, 683, 693
Lin, Yu-Xuan, 768
Oma, Ryuji, 612
Liu, Te-Lung, 441
Ozaki, Ryo, 362
Liu, Wen-chao, 93
Ozaki, Tomonobu, 305, 314
Liu, Yi, 234, 422
Ozera, Kosuke, 234
Liu, Yu-Jia, 14
Long, Wen-feng, 14
Luo, Yu, 1007 P
Pacini, Alessandro, 724
M Pagnotta, Fabio, 745
Maabreh, Majdi, 897 Pang, Jiayi, 561
Maeda, Hiroshi, 275 Papageorgiou, Nikos, 985
Maisto, Alessandro, 805 Park, Hyunhee, 256, 433
Majeed Khan, Abdul Basit, 1097 Park, Mirang, 683, 693
Mambretti, Joe, 441 Park, Seunghyun, 433
Mansourbeigi, Seyed M.-H., 623 Patiniotakis, Ioannis, 996
Marcantoni, Fausto, 724 Pelosi, Serena, 805
Marchang, Ningrinla, 702 Piermarteri, Andrea, 724
Materka, Katarzyna, 1041 Polito, Massimiliano, 805
Matsui, Kanae, 295 Pradhan, Dipesh, 1030
Matsuo, Keita, 422, 604 Prakash, Satya, 129
Mazzara, Manuel, 116 Przeździȩk, Tomasz, 1041
Mehmood, Tahir, 1145 Pujolle, Guy, 662
Meng, Xiang Zheng, 275
Mentzas, Gregoris, 985, 996 Q
Micheletti, Matteo, 724 Qasim, Umar, 371
Minieri, Maurizio, 860 Qureshi, Talha Naeem, 1084
1172 Author Index